chore(workspace): add ourios-core skeleton by jensholdgaard · Pull Request #3 · jensholdgaard/ourios

jensholdgaard · 2026-04-28T13:13:17Z

Summary

Lights up the first concrete workspace member per CLAUDE.md §7.

Cargo.toml — crates/ourios-core added to members; comment block reframed as the remaining list.
crates/ourios-core/Cargo.toml — package metadata + [lints] workspace = true (the workspace already declares unsafe_code = "deny" and clippy::pedantic = "warn").
crates/ourios-core/src/lib.rs — one line: //! Foundational types for Ourios.

Foundational types (Tenant newtype, LogRecord, error enum) deliberately do not land in this PR. Per CLAUDE.md §3.7 Multi-tenancy is not bolted on, those types touch a §3 invariant and per §5.1 want an RFC first. This PR is mechanically about getting the workspace to a buildable state so the cargo CI jobs (fmt, clippy, test) start exercising real code instead of an empty members = [] workspace.

Invariants and hazards touched

None directly. The crate has no implementation yet. Subsequent PRs adding Tenant etc. will touch CLAUDE.md §3.7 and require an RFC.

Test plan

mdbook build — clean (verified locally).
cargo fmt --all --check — not run locally (toolchain not installed on author's machine; pinned in rust-toolchain.toml). CI is the gate.
cargo clippy --all-targets --all-features -- -D warnings — same, CI is the gate.
cargo test --all-features — same, CI is the gate.
CI Success aggregate check passes.

For a lib.rs consisting of a single doc-comment line, all three cargo checks are trivially correct, but per CLAUDE.md §6.6 ("If a check is not set up yet, say so explicitly instead of claiming green") I'm flagging that local verification was incomplete and CI is authoritative for this PR.

🤖 Generated with Claude Code

Lights up the first concrete workspace member per CLAUDE.md §7. The crate carries only a single-line module doc; foundational types (Tenant, log record, errors) land via subsequent RFC-driven PRs that operationalise §3.7 (multi-tenancy is not bolted on) and the §3 invariants more broadly. The shape: - Cargo.toml — workspace member added; comment reframed as the remaining list, since one is now landed. - crates/ourios-core/Cargo.toml — package metadata + workspace lints. - crates/ourios-core/src/lib.rs — one line: `//! Foundational types for Ourios.` Local verification: - mdbook build: clean. - cargo fmt/clippy/test: not run locally (toolchain not installed on author's machine; rustup-pinned in rust-toolchain.toml is the authoritative source). For an empty lib.rs all three are trivially correct; CI is the real gate. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

jensholdgaard · 2026-04-28T13:23:35Z

Local verification now complete (rustup installed after the PR was opened):

✅ cargo fmt --all --check
✅ cargo clippy --all-targets --all-features -- -D warnings — zero warnings
✅ cargo test --all-features — 0 passed, 0 failed (empty crate, expected)
✅ mdbook build

Toolchain: cargo 1.95.0 / rustc 1.95.0, matching rust-toolchain.toml's stable channel.

Copilot

Pull request overview

Adds an initial ourios-core crate to the workspace so CI jobs (fmt, clippy, test) run against a non-empty workspace, per CLAUDE.md layout guidance.

Changes:

Add crates/ourios-core as the first concrete [workspace] member.
Introduce crates/ourios-core crate manifest with workspace-inherited metadata and lints.
Add a minimal lib.rs with crate-level documentation.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
Cargo.toml	Adds `crates/ourios-core` to workspace members and updates the “remaining members” comment block.
crates/ourios-core/Cargo.toml	Defines new crate metadata and enables workspace lints.
crates/ourios-core/src/lib.rs	Adds minimal crate-level doc comment to make the crate buildable.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Signed-off-by: Jens Holdgaard Pedersen <jens@holdgaard.org>

* feat(querier): scaffold ourios-querier crate — RFC 0007 red gate New crate (pillar #3, the read path). Public API surface from RFC 0007 §4.1: Querier, QueryRequest (tenant + time bounds; the RFC 0002 parsed-query field is deferred), QueryResult/QueryStats (row_groups_scanned/pruned + bytes_read — the fields B1 asserts on), and a QueryError enum that leaks no datafusion/arrow/SQL types (hazard §4.6). Querier::run is unimplemented!() (red gate). 5 #[ignore]'d acceptance stubs (tests/acceptance.rs) map RFC0007.1 (B1 pushdown prunes), .2 (B2 latency scales with result not corpus), .3 (no DataFusion/SQL leakage), .4 (forward-compat reads), .5 (tenant isolation). Added to the workspace members. Query EXECUTION is deferred: the DSL→LogicalPlan lowering needs RFC 0002's undecided Branch A/B. RFC 0007 scopes this layer as branch-independent, so the surface + B1/B2 criteria land now; datafusion + ourios-parquet deps arrive with the execution slice (kept out of the scaffold to keep the red-gate build minimal). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fixup! feat(querier): scaffold ourios-querier crate — RFC 0007 red gate * fixup! feat(querier): scaffold ourios-querier crate — RFC 0007 red gate --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…(RFC0008.8) Review (Copilot + self-review) surfaced that making ingest concurrent broke the 'miner ingests in WAL-append order' property the serialized pre-.8 pipeline guaranteed — template ids are assigned first-seen, so an out-of-order live tree wouldn't match a WAL-order replay (snapshot-restore §3.5.3) and the rotation hook could snapshot past its claimed high-water (double-apply). Maintainer chose the ordered hand-off. - commit() returns CommitOutcome { seq, result }: the append sequence + the durability result. After durability the pipeline drives an in-order gate (await_ingest_turn(seq) → miner work → complete_ingest(seq)), so the miner ingests in exact seq (= WAL-append) order while fsyncs still batch. complete_ingest runs for every seq-bearing outcome (success or sync failure) so the gate never stalls; only successful commits advance last_durable, so the snapshot high-water never passes a failed sync (its tail replay re-covers those frames — no §3.5.3 divergence). Fixes the rotation-hook ordering Copilot flagged (pipeline.rs): before/now and the hook now run in-order under the gate. - arm_flush: a segment-fill cut now wakes a pending flush via a Notify instead of spawning a flush per fill-crossing append (Copilot — at most one extra flush task per window, not one per append). - Tests: a new concurrent-ingest order test (24 distinct templates → live miner equals a WAL-order replay control; diverges without the gate); latency test reworked to the CI-robust scaling-ordering proof (the absolute ±30% band flaked on a shared runner where fixed per-flush overhead dominates small windows — the scaling ordering is the faithful, robust proof that latency tracks the window). The inline append stays on the async task (Copilot #3): it's a fast page-cache write under the single-writer journal lock, not the fsync (which is offloaded) — offloading a lock-guarded microsecond write per request would add more scheduling overhead than it saves. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(ingester): windowed group-commit coordinator for batched fsync (RFC0008.8) The WAL stays single-writer + synchronous (§6.1 makes batching the caller's job); a new CommitCoordinator in ourios-ingester folds N concurrent per-request appends into one fsync per wal_batch_window_ms window (or an early cut when unflushed bytes reach the segment size), preserving §3.4: a commit returns Ok only after a sync that covered its frame returned Ok. Coverage is by a monotonic append sequence number, not the byte offset — a sync makes durable everything appended before it, so a waiter is covered exactly when seq <= the flush's covered_seq (its bytes were in the file before that sync). The journal Mutex is never held across an await (append takes it; the blocking sync re-takes it inside spawn_blocking). Outcomes broadcast over a tokio watch, applied MONOTONICALLY in covered_seq: two flushes can be in flight (a fill-cut spawned while a window flush is pending) and broadcast out of order, so a stale lower-covered_seq outcome must never overwrite a higher one (it would strand a covered waiter under idle-after-burst traffic) — dropping it is also §3.4-safe (a higher failed flush masking a lower success only yields a spurious error→retry, never an ack of a non-durable frame). IngestPipeline now owns Arc<CommitCoordinator> + Mutex<MinerCluster>; ingest() is async (&self) and SharedPipeline = Arc<IngestPipeline>, so concurrent requests batch instead of serializing end-to-end. Both transports await ingest() directly (the blocking fsync moved into the coordinator's spawn_blocking). Rotation hook + last_durable seeding preserved. Journal::sync now returns WalOffset (was Option) + unflushed_bytes(); the No-offset unit test is replaced by the seeding-and-supersede contract it still asserts (surfaced per §6.2 — internal trait, the real WAL always has an offset). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test: flip RFC0008.8 batched-fsync, relocated to ourios-ingester Ack latency is a receiver concept, so the acceptance test moves from ourios-wal/tests to ourios-ingester/tests (relocation precedent: RFC0001.5/.6 miner→querier). Three arms: P99 ack latency tracks the batch window (and scales p99(10)<p99(50)<p99(150) — the noise-robust assertion; windows scaled to {10,50,150}ms to keep the test ~2s, documented in-file); syncs advance per-batch not per-record (appends_per_sync >> 1); and the §3.4 gate (a commit Ok implies a covering successful sync, via a recording spy). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(ingester): ordered miner hand-off + monotonic-bound review fixes (RFC0008.8) Review (Copilot + self-review) surfaced that making ingest concurrent broke the 'miner ingests in WAL-append order' property the serialized pre-.8 pipeline guaranteed — template ids are assigned first-seen, so an out-of-order live tree wouldn't match a WAL-order replay (snapshot-restore §3.5.3) and the rotation hook could snapshot past its claimed high-water (double-apply). Maintainer chose the ordered hand-off. - commit() returns CommitOutcome { seq, result }: the append sequence + the durability result. After durability the pipeline drives an in-order gate (await_ingest_turn(seq) → miner work → complete_ingest(seq)), so the miner ingests in exact seq (= WAL-append) order while fsyncs still batch. complete_ingest runs for every seq-bearing outcome (success or sync failure) so the gate never stalls; only successful commits advance last_durable, so the snapshot high-water never passes a failed sync (its tail replay re-covers those frames — no §3.5.3 divergence). Fixes the rotation-hook ordering Copilot flagged (pipeline.rs): before/now and the hook now run in-order under the gate. - arm_flush: a segment-fill cut now wakes a pending flush via a Notify instead of spawning a flush per fill-crossing append (Copilot — at most one extra flush task per window, not one per append). - Tests: a new concurrent-ingest order test (24 distinct templates → live miner equals a WAL-order replay control; diverges without the gate); latency test reworked to the CI-robust scaling-ordering proof (the absolute ±30% band flaked on a shared runner where fixed per-flush overhead dominates small windows — the scaling ordering is the faithful, robust proof that latency tracks the window). The inline append stays on the async task (Copilot #3): it's a fast page-cache write under the single-writer journal lock, not the fsync (which is offloaded) — offloading a lock-guarded microsecond write per request would add more scheduling overhead than it saves. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(ingester): panic-safety on the ingest path + non-accumulating fill signal Review round 2 (all real): - Gate deadlock on panic: a panic between await_ingest_turn(seq) and the release would leave the in-order gate stuck at seq and deadlock every later ingest (the miner has expect() invariants). An IngestGateGuard now releases the hand-off on Drop, so the gate advances even during unwinding. - Transport no-panic contract: making ingest a direct .await dropped the spawn_blocking+JoinError panel that contained panics. Both handlers now run ingest on a tokio::spawn task and map a JoinError to INTERNAL/500, so a pipeline/miner panic is a controlled response, not a dropped connection (spawn, not spawn_blocking — ingest is async). - Fill signal: Notify stores permits, so repeated fill-cuts could linger and spuriously cut a later window short. Replaced with a watch counter the armed flush task subscribes to at spawn — a bump only wakes the task subscribed before it, no cross-cycle carryover. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…6) (#245) * feat(server): wire the RFC 0014 data write path live (rfc0013 green .6) The production data path now flows OTLP → WAL → miner → ParquetRecordSink → object store. The miner emits through a SharedParquetSink (a cloneable Arc<Mutex<ParquetRecordSink>> handle); the pipeline drives the flush triggers the sink can't observe itself — flush_all on WAL segment rotation (RFC0014.3) and flush_aged on a batch-window age sweep (RFC0014.2). The store is a LocalFileSystem-backed Store rooted at bucket_root; S3 selection (RFC 0004) is the RFC 0014 §7 follow-on. WAL-durability invariant (CLAUDE.md §3.4, hazard #3). The sink buffer is an in-memory accelerator, never the durability of record — records are WAL-fsync'd before they reach it. No-loss across a crash is preserved by a single ordering rule applied at every miner-snapshot cadence point (post-recovery, rotation, graceful shutdown): flush the sink *before* writing the snapshot. That keeps the miner's snapshot horizon at or below the sink's flushed horizon, so startup recovery's miner-gated replay re-emits every un-flushed acknowledged record into a fresh sink (the crash discards only the volatile buffer). At-least-once: records flushed just before a crash may be re-flushed on restart; no record is lost. Multi-tenancy (§3.7) is unchanged — buffers are keyed by PartitionKey, which carries tenant_id. RFC0013.6 (WAL stays local) is greened end to end through the served binary in ourios-server/tests/rfc0013_6_wal_stays_local.rs: it ingests over HTTP, SIGTERMs (the shutdown drain flushes), and asserts only Parquet/manifest objects land under bucket_root while the WAL *.wal segments stay on the disjoint local wal_root. The stub in ourios-parquet/tests is redirected there (that crate has no WAL or server to observe). All eight RFC 0013 §5 scenarios now pass → RFC 0013 green. RFC0014.5 (crash no-loss) stays #[ignore]d pending its dedicated crash fixture (the next PR), which exercises this same recovery path under a real SIGKILL. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): offload age sweep to the blocking pool; create the store root Addresses Copilot review on #245: - `spawn_age_sweep` ran `flush_aged` (Parquet encode + blocking store I/O) directly on a runtime worker; move it to `spawn_blocking` so periodic sweeps never stall the receiver (esp. against S3). - `Store::local` errors if `bucket_root` is missing, which regressed the server's ability to start on a not-yet-created `OURIOS_BUCKET_ROOT`; `create_dir_all` the root before opening the store. The RFC0013.6 test now omits its own `create_dir_all`, so it covers this path. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): gate the cadence snapshot on the sink actually draining Closes a no-loss hole Copilot flagged on #245. `flush_all` retains a partition whose store write fails (the WAL is the durability of record), but the cadence points wrote the miner snapshot unconditionally afterward. With the store unavailable that advances the snapshot horizon past records that never reached object storage — and recovery suppresses frames at or below the horizon, so on the next start they are never re-emitted into a fresh sink. That is data loss precisely when the store is down. Fix: a `flush_then_snapshot` helper flushes, then writes the snapshot only if `buffered_records() == 0`. If the sink didn't drain it skips the snapshot (logged), so the horizon can't outrun the flushed horizon and recovery re-mines the un-flushed records. Applied at all three cadence points (post-recovery, rotation, shutdown). The rotation hook — which runs on the request path — wraps the blocking flush + snapshot in `block_in_place` so it doesn't stall a runtime worker. Tested: a fault-injection unit test replaces the store root with a file so `put_blocking` fails, then asserts the snapshot is skipped and the records stay buffered (not lost); plus the drained happy path. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): steady age-sweep cadence; block_in_place all cadence flushes Addresses Copilot review round 3 on #245: - Age sweep: set the interval's missed-tick behavior to `Delay` so a slow sweep (S3) can't trigger catch-up bursts of back-to-back flushes. - `flush_then_snapshot` doc: the bool is "did the sink drain" — clarified that a `write_snapshots` failure still returns `true` (the data reached the store; the snapshot is a rebuildable-cache miss), resolving the contradiction with the old "snapshot written" wording. - Wrap the post-recovery and shutdown cadence flushes in `block_in_place` too (not just the rotation hook), so their blocking Parquet/store I/O doesn't stall a runtime worker. - `SharedParquetSink` doc: correct the "short critical sections" claim — `flush_all`/`flush_aged` hold the mutex across encode + `put_blocking`, so they are blocking sections (benign sub-ms on the local backend). Moving the encode+put outside the lock is tracked for the S3 backend (RFC 0014 §7 / RFC 0013), where PUTs are slow enough to matter. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): await the age-sweep on shutdown; test-helper clarity Addresses review round 4 on #245: - Shutdown awaits the age-sweep task instead of aborting it. Aborting the async task would orphan an in-flight `spawn_blocking` flush still holding the sink mutex (which the shutdown drain would then wait on anyway); the task already observes the `shutdown` watch signal, so awaiting lets the in-flight flush finish and the task exit cleanly (Copilot). - RFC0013.6 test: `http_post_logs` `.expect()`s the flush, and `files_under` fails fast on `read_dir`/entry errors rather than silently skipping — in a controlled temp tree those are real problems (CodeRabbit). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): soften the snapshot-write-failure log wording Round 5 (Copilot): a failed snapshot write doesn't force a *full* replay — prior snapshots may survive and partial per-tenant writes may have landed. Reword the cadence log from "next start full-replays" to "next start may replay more from the WAL" so it doesn't mislead operators. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

…302) Round 3 on #312 — close the empty-body window fully: a mined record must never become query-visible before its template's audit event is durable, under concurrency and the inline size trigger (strengthens `CLAUDE.md` §3.3 to hold on every publication path). #4 — `emit` no longer blocks behind flush I/O. The audit sink's flush now drains the buffer under the lock, releases it, does the `AuditWriter` store I/O **unlocked**, then re-locks only to settle counters and requeue a transient failure's events ahead of anything `emit` buffered meanwhile (mirrors the record sink's documented "drain under the lock, I/O unlocked, re-lock to settle"). A slow flush can't stall the request path. #3 — the buffer is hard-bounded. The soft ceiling still signals an eager off-runtime flush; a new hard cap (`AUDIT_SINK_MAX_EVENTS`, well above the ceiling) is the OOM backstop: at the cap `emit` drops (counted via the new `ourios.audit_sink.dropped` metric, logged once) rather than grow without bound under sustained store-unavailability. Dropped template events degrade those templates to retained/empty bodies until the WAL re-mines them on restart — bounded memory is the deliberate trade. #1/#2 — publication is audit-ordered and race-free via snapshot-then- ordered-write. A new `PublishCoordinator` (ourios-ingester) drains both sink buffers into owned batches under the pipeline's miner lock (a microsecond memory move, no I/O — atomic w.r.t. `ingest`, closing the cadence TOCTOU race), then writes off-lock: the audit batch to durability first, the record partitions only after. A transient audit failure holds the records (requeued, retried next cadence); a permanent audit failure drops the audit batch and still publishes the records (the documented degraded case). The receiver's age-sweep now publishes through it. The inline size/ceiling trigger routes through a new record-sink audit barrier (`ParquetRecordSink::with_audit_barrier`) that flushes the audit sink to durability before the partition is put — race-free because that publish runs under the miner lock. Rotation/shutdown already drain audit-before- record under the miner lock (`flush_then_snapshot`), unchanged. The record sink gains a drain/publish/requeue split (`drain_aged` / `drain_all` / `requeue` / `publish_owned`) so the coordinator can move the encode+put off the lock; its existing RFC 0014 emit/flush behavior and tests are intact. Tests: the coordinator holds records when the audit write fails transiently (no data partition published though the data store is healthy); the size trigger flushes audit-before-publish and is skipped when audit can't drain; the hard cap drops + bounds; transient retains vs permanent drops; the metrics export; plus all round-1/2 tests and the #302 regression stay green. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

) (#312) * fix(server): wire a buffering audit sink into the receiver (#302) The OTLP receiver wired the template miner with a record sink but no audit sink, so the miner's `template_created` / `template_widened` / `template_type_expanded` events never reached the RFC 0005 audit Parquet stream. The querier's read-time registry (RFC 0017 `derive_template_registry`) was therefore empty and `render_log_body` fell back to the row's retained `body` — empty for clean, high-confidence rows — so queries over freshly-ingested clean logs returned empty body text, breaking `CLAUDE.md` §3.3. Mirror the RFC 0014 record sink rather than wiring `ParquetAuditSink` directly (which does a blocking per-event store write — request-path stall + one tiny file per event, hazard #4): - `BufferingAuditSink` / `SharedParquetAuditSink` (ourios-ingester): `emit` buffers cheaply on the request path; `flush` drains the buffer, groups events by audit partition, and writes each partition's batch with one `AuditWriter` (open_in → append_events → close) — few files, not one-per-event. A failed partition write retains its events (the WAL is the durability of record); an empty-buffer flush is a no-op. - The receiver constructs the sink on the same `Store`, wires it via `MinerCluster::with_audit_sink(...).with_record_sink(...)` before recovery (so replay re-emits template events), and flushes it off the async runtime at the same cadence + rotation + shutdown points as the record sink — audit *before* records (durable no later than the rows it describes), with the snapshot gated on both sinks draining. - Expose `derive_audit_partition` from ourios-parquet for the grouping. Tests: unit tests for the buffering sink (per-partition batching round trip + empty-buffer no-op); an in-process receiver test that ingests clean logs, drains, derives the registry, and asserts every clean row reconstructs `Faithful` from its template rather than empty. The RFC0019 `.3`/`.5` localstack scenarios now also assert the returned body text. `rfc0013_6` scopes its data round-trip to `data/` so the new `audit/` files aren't read with the data schema. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(parquet): drop intra-doc link to a private item derive_audit_partition is now pub; its doc referenced the private audit_partition_matches via an intra-doc link, which trips rustdoc::private-intra-doc-links under cargo doc -D warnings. Use plain backticks (the recurring private-item-doc convention). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): audit-sink metrics, error classification, bounded buffer (#302) Address PR #312 review feedback on the receiver audit sink. OTel metrics (§6.3): add `ourios.audit_sink.*` instruments mirroring the record sink's `SinkMetrics` — `buffer.usage` (observable gauge of buffered events), `flushes`, `flush.events`, `flush.errors` (split transient vs permanent via the new `ourios.audit_sink.flush.outcome` attribute), and `derive.errors`. Names go through the weaver registry (`semconv/registry/{metrics,attributes}.yaml`); the generated `ourios-semconv` constants are regenerated, not hand-written. Instruments resolve through the global meter (no-op without a provider). A dedicated test binary asserts the stream exports (separate process — `init_in_memory` installs the global provider). Data integrity: classify a failed partition flush. A store-`Io` error is transient → retain + retry (the WAL is the durability of record); a `Batch` / `Parquet` / `PartitionMismatch` / `Poisoned` error is permanent → drop + count, so one malformed event can't requeue forever and wedge every newer good event for that tenant/day behind it. §3.3 flush gating: in both the age-sweep and `flush_then_snapshot`, flush the audit sink first and skip the record flush this cycle if it didn't fully drain — a non-empty buffer means a transient store error (permanents drop), so the record flush to the same store would fail anyway, and flushing it would expose a clean row before its template event is durable. Bounded buffer: `emit` stays non-blocking but enforces a soft event ceiling (default 100k); reaching it signals a `tokio::sync::Notify` the age-sweep selects on, so adversarial template churn flushes promptly off the runtime rather than growing the buffer until OOM. Signal-to-flush, never drop. Tests: poison-pill (permanent drops + counts, does not requeue; transient retains), flush-gating (record flush skipped while audit retains), bounding (emit past the ceiling fires the notify), plus the metrics-export test. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(server): audit-ordered publication, non-blocking flush, hard cap (#302) Round 3 on #312 — close the empty-body window fully: a mined record must never become query-visible before its template's audit event is durable, under concurrency and the inline size trigger (strengthens `CLAUDE.md` §3.3 to hold on every publication path). #4 — `emit` no longer blocks behind flush I/O. The audit sink's flush now drains the buffer under the lock, releases it, does the `AuditWriter` store I/O **unlocked**, then re-locks only to settle counters and requeue a transient failure's events ahead of anything `emit` buffered meanwhile (mirrors the record sink's documented "drain under the lock, I/O unlocked, re-lock to settle"). A slow flush can't stall the request path. #3 — the buffer is hard-bounded. The soft ceiling still signals an eager off-runtime flush; a new hard cap (`AUDIT_SINK_MAX_EVENTS`, well above the ceiling) is the OOM backstop: at the cap `emit` drops (counted via the new `ourios.audit_sink.dropped` metric, logged once) rather than grow without bound under sustained store-unavailability. Dropped template events degrade those templates to retained/empty bodies until the WAL re-mines them on restart — bounded memory is the deliberate trade. #1/#2 — publication is audit-ordered and race-free via snapshot-then- ordered-write. A new `PublishCoordinator` (ourios-ingester) drains both sink buffers into owned batches under the pipeline's miner lock (a microsecond memory move, no I/O — atomic w.r.t. `ingest`, closing the cadence TOCTOU race), then writes off-lock: the audit batch to durability first, the record partitions only after. A transient audit failure holds the records (requeued, retried next cadence); a permanent audit failure drops the audit batch and still publishes the records (the documented degraded case). The receiver's age-sweep now publishes through it. The inline size/ceiling trigger routes through a new record-sink audit barrier (`ParquetRecordSink::with_audit_barrier`) that flushes the audit sink to durability before the partition is put — race-free because that publish runs under the miner lock. Rotation/shutdown already drain audit-before- record under the miner lock (`flush_then_snapshot`), unchanged. The record sink gains a drain/publish/requeue split (`drain_aged` / `drain_all` / `requeue` / `publish_owned`) so the coordinator can move the encode+put off the lock; its existing RFC 0014 emit/flush behavior and tests are intact. Tests: the coordinator holds records when the audit write fails transiently (no data partition published though the data store is healthy); the size trigger flushes audit-before-publish and is skipped when audit can't drain; the hard cap drops + bounds; transient retains vs permanent drops; the metrics export; plus all round-1/2 tests and the #302 regression stay green. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(ingester): keep requeued partitions aged for prompt retry PublishCoordinator.requeue re-buffers a transient-failed batch ahead of records emit added during the off-lock publish, but left PartitionBuffer.oldest at the newer records' timestamp — so the already-aged requeued records could miss the next age-sweep and retry late. Pin oldest to the age threshold (min with the existing oldest) on requeue so the next sweep re-drains promptly. Test: requeue_keeps_the_partition_aged_for_prompt_retry. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(ingester): audit buffer retains, never drops (reverse the hard cap) (#302) Round 5 on #312. Item 1: reword the audit_sink module-doc typo "The Drain miner emits…" → "The template miner emits…". Item 2 (CodeRabbit Major — the round-2 hard-cap drop could lose data): the hard cap dropped audit events at `AUDIT_SINK_MAX_EVENTS`, which is unsafe. A dropped event isn't counted by `buffered_events()`, so the no-loss snapshot gate (`flush_then_snapshot`) doesn't see it, the miner snapshot advances past that line's WAL position, and on restart the template event is never re-mined → those clean rows become permanently unreconstructable (a §3.3 violation), not merely degraded-until-restart. Adopt the record sink's posture (follow the reference, §5.4): under sustained store-unavailability the audit buffer is RETAINED and may transiently exceed the ceiling — never dropped. The WAL is the durability of record; the snapshot gate prevents loss because it won't advance while the buffer is non-empty. - Delete `AUDIT_SINK_MAX_EVENTS` and the drop branch in `buffer_event`; `emit` always buffers. The soft ceiling still fires the `Notify` for an eager off-runtime flush — the bound for the realistic (healthy-store) case. `requeue_ahead` no longer caps/drops. - Remove the `ourios.audit_sink.dropped` metric: reverted its `semconv/registry/metrics.yaml` entry, regenerated `ourios-semconv` (the const is gone; weaver no-diff verified), and dropped its use. - Replace the `hard_cap_drops_and_bounds_the_buffer` test with `persistent_store_failure_retains_every_event_never_drops`: under a persistently failing store, repeated emit + flush retains every event (buffer grows past the ceiling) and drops nothing. - Module docs state the posture explicitly (healthy store → ceiling + Notify bound it; sustained outage → retained, like the record sink; OOM under a total outage is the same accepted failure mode the record sink carries). The transient-vs-permanent flush classification (Io retain / Batch + Parquet + PartitionMismatch + Poisoned drop+count) is unchanged — that's about un-writable content, unrelated to the memory bound. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

The fn doc referenced the old ≈3 MiB figure and the constant's doc didn't flag 3 MiB as the PREVIOUS value; both now point at FLUSH_BYTES and date the run #3 measurement to the prior cap. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…y) (#477) * fix(bench): 1.5 MB flush cap — Loki inflates OTLP internally (run #3) Run #3 (29164195463) failed at the same corpus region as run #2 with the IDENTICAL internal size (5,276,869 bytes) — but this time our fail-fast did NOT fire, proving our HTTP payload was under 4 MiB while Loki's internal gRPC message exceeded it. Loki's OTLP→logproto translation INFLATES content ≥1.76x here: OTLP shares resource/scope attributes per batch; the internal push repeats labels and structured metadata per entry. Drop FLUSH_BYTES to 1.5 MB (≥2.6x inflation headroom under the stock 4 MiB cap). push_otlp keeps asserting our own encoded size as a floor guarantee — if a single mega-record alone ever exceeds the cap, the fail-fast will name it precisely. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(bench): sync pusher docs to the 1.5 MB cap (Copilot review) The fn doc referenced the old ≈3 MiB figure and the constant's doc didn't flag 3 MiB as the PREVIOUS value; both now point at FLUSH_BYTES and date the run #3 measurement to the prior cap. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…flation) (#478) * fix(bench): raise Loki's internal gRPC cap — single-line inflation (run #4) Run #4 (29165198664) failed on the SAME ~5.27 MB internal message as runs #2/#3 despite the outer cap halving (3 MiB → 1.5 MB), and the fail-fast stayed silent — decisive: a single kafka LogsData line's content alone inflates past Loki's stock 4 MiB internal gRPC cap. No outer batching can split an indivisible unit. Add -server.grpc-server-max-recv/send-msg-size=16 MiB to the indicative run's documented ingest-side flags (standard operator tuning, in Loki's favour — it lets Loki accept the data at all). This preserves the identical-ingest precondition the equivalence check requires; skipping the line would silently unequalize the two corpora. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(bench): correct dskit flag names for the gRPC msg-size cap Copilot caught that the flags are -server.grpc-max-recv/send-msg-size- bytes (dskit's server registry, defaults exactly the 4 MiB we hit), not -server.grpc-server-max-*. The wrong names would have failed Loki's startup and burned run #5. Verified against dskit source. Also backtick the kafka service name in the comment. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…u32::MAX bound PR review findings #1 and #3. The structured JSON surface's group_term accepted {resource|attr} field objects in a by-list, but the string DSL's group_term = field production (§7 v1.1) is bare-field-only — so the structured surface could express count/aggregate-by queries the string grammar cannot, which then failed in planning instead of at validation. RawGroupTerm::into_ir now rejects a {resource|attr} object with a clean DslError, and the schema gains a bare_field $defs entry so schema validation itself rejects the shape instead of only the runtime converter. The schema's param integer also gets an explicit maximum (u32::MAX) so an out-of-range param slot fails schema validation cleanly rather than succeeding the schema and then failing Rust deserialization. Adds schema instance-list cases (resource/attr group term, param past u32::MAX) and a structured.rs unit test for the runtime rejection. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…bucket(w) (#533) * feat(querier): rfc 0002 green — count-by execution with param(n) and bucket(w) The aggregation-execution slice of the RFC 0002 amendment 2026-07-15 (RFC 0031 L4): `count [by …]` now executes end-to-end, discharging §5 scenarios RFC0002.12/.13/.15/.16. RFC0002.14 (the grammar/compile error contracts) stays an ignored red stub for its own slice. Surface (§7 v1.1 / §6.4 amendment): - IR: the aggregation `by`-list widens from `Vec<Field>` to `Vec<GroupTerm>` (field | `param(n)` | `bucket(duration)`). - Parser: `group_list`/`group_term` productions, confined to `by`-lists; positive + negative parse tests per production. - Structured surface: `{"param": n}` / `{"bucket": "<duration>"}` by-elements (widths validated by the string-DSL lexer, RFC0002.2); `structured_query.schema.json` gains the additive `group_term` def (snapshot-gated by RFC0002.11, which also gains instances). - Serializer: group terms round-trip (corpus + proptest generator). Compile (§6.3/§6.5 amendment): - `compile::validate` lifts the `count` rejection ONLY — sum/min/max/avg, sort, project, render keep the explicit rejection. Enforces the single-template pinning rule for `param(n)` (top-conjunctive `template_id == N`, all naming one N; `resolves_to` does not pin), positive bucket widths, and the duplicate-term rules. - Group terms lower as expressions inside the existing Aggregate row: `param(n)` = `array_element(params, n+1).value` (stored string form, no type promotion); `bucket(w)` = floor division of the effective timestamp (with the §3.9 `time_unix_nano` fallback) into half-open epoch-aligned UTC windows; `service` = the RFC 0022 promoted column. Execute: - One grouped-count scan per aggregation query (Filter → Aggregate, the drift precedent) with a row-level `tenant_id` guard mirroring drift's (§3.7 — group values are row contents). `rows` stays the total matching count, derived from the same scan. - Short/NULL `param(n)` rows are EXCLUDED from every group (no synthetic absent key) and tallied on the new `QueryStats.rows_excluded`, surfaced on the RFC 0016 stats DTO (RFC0002.15). - Result carrier: `QueryResult.aggregate: Option<Vec<AggregateGroup>>` (`key: Vec<String>` per by-term in query order — bucket keys RFC 3339 UTC window starts — sorted, engine-free per hazard §4.6); the RFC 0016 response gains the additive `aggregate` field so the HTTP surface cannot silently drop the map. - RFC0002.16 honest bytes: the total is the group-column scan alone — zero row materialization, zero template-map acquisition (the RFC 0033 acquisition was already lazy; the aggregation path never renders). Invariants: hazard §4.6 (no DataFusion/arrow/SQL crosses the surface — plain strings/ints only); §3.7 multi-tenancy (partition scope + the new row-level tenant filter on the aggregation plan). Contract changes sanctioned by the maintainer-merged amendment (#531): the `count` case moves out of rfc0002_6_unsupported_stage_rejected (RFC0002.12 names the lift), and rfc0005_14's error-precedence probe switches from `count` to the still-rejected `render`. Verified: cargo fmt --check; workspace clippy --all-targets --all-features -D warnings; strict rustdoc (ourios-querier); full cargo nextest run (1107 passed); .12/.13/.15/.16 force-run green, .14 still ignored-failing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(querier): structured by-list rejects resource/attr paths + param u32::MAX bound PR review findings #1 and #3. The structured JSON surface's group_term accepted {resource|attr} field objects in a by-list, but the string DSL's group_term = field production (§7 v1.1) is bare-field-only — so the structured surface could express count/aggregate-by queries the string grammar cannot, which then failed in planning instead of at validation. RawGroupTerm::into_ir now rejects a {resource|attr} object with a clean DslError, and the schema gains a bare_field $defs entry so schema validation itself rejects the shape instead of only the runtime converter. The schema's param integer also gets an explicit maximum (u32::MAX) so an out-of-range param slot fails schema validation cleanly rather than succeeding the schema and then failing Rust deserialization. Adds schema instance-list cases (resource/attr group term, param past u32::MAX) and a structured.rs unit test for the runtime rejection. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(querier): reject count+limit; property-test the §6.3 planner invariants PR review findings #2 and #4. #2: `count [by …] | limit n` silently dropped the `limit` — execution terminates in `Terminal::Aggregate`, which never consults `plan.limit` (the aggregation map is the whole result; group-limiting semantics aren't implemented). `validate()` now rejects the combination with a clear QueryError::InvalidQuery instead of quietly returning the wrong thing. #4: pin detection (top-conjunctive `template_id == N`; `or`/`not`/ `resolves_to` don't pin), param-position duplication (at most one `param(n)` per distinct n), and bucket constraints (positive width, at most one `bucket(...)`) were covered only by hand-picked examples. Adds a proptest generating arbitrary predicates and by-lists, checked against an independently tracked ground truth (ground truth recorded alongside generation, not derived from the code under test), covering both `pinned_template_id` and `validate()`'s accept/reject decision. The hand-picked examples stay as-is (CLAUDE.md §6.2). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(querier): event_name in group-term generator; tenant + NULL-param regressions PR review findings #5, #6, #7. #5: the RFC0002.7 round-trip generator's `bare_field()` — shared by `path_field()`, `group_term()`, and the `project` field list — omitted `Field::EventName`, so grouped-query round-trips never covered `count by event_name`. It's a valid bare field everywhere the real grammar's `bare_field` (parse.rs) allows it, so fixed in place. #6: `rfc0002_12_count_by_matches_naive_oracle`'s foreign-tenant fixture is written via `write_all`, which partitions by the record's own `tenant_id` — so the "b" row lands under tenant "b"'s own directory and the row-level `tenant_id == tenant` backstop in `execute_aggregate` (CLAUDE.md §3.7) is never exercised, only directory-level scoping. Adds `rfc0002_12_aggregation_tenant_backstop_excludes_misplaced_row`, which plants a tenant "b" row *inside* tenant "a"'s partition directory (the shape a partitioning bug or on-disk corruption would produce — the `ourios-parquet` writer's RFC 0005 §3.9 row-vs-path contract refuses a mismatched tenant_id at write time, so the row is written honestly then relocated) and asserts the backstop filter, not partitioning, keeps it out of both the count and the group map. Manually verified this test fails without the backstop filter, confirming it exercises the guard. #7: RFC0002.15 covered a `params` list shorter than `n + 1`, but not the distinct case of a list that HAS slot n whose own `value` decodes as Parquet-level NULL (the field is nullable — RFC 0005 §3.2 — even though `Param.value` is a non-`Option` Rust `String`, so only a raw/corrupted writer can produce it). Adds `rfc0002_15_present_but_null_param_slot_excluded_and_tallied`, built with a raw arrow-array batch (mirroring `forward_compat.rs`'s schema-drift fixtures) so the disposition is proven on the actual `decode_aggregate` code path rather than assumed from the short-list case. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(querier): rfc 0002 — non_exhaustive QueryStats, typed group-null literal QueryStats gains #[non_exhaustive] matching QueryResult's convention. The rows_excluded doc comments now scope to any NULL group key, not just param(n). The absent-OPTIONAL-column NULL substitute in the aggregate group-term compiler now carries the field's real Arrow type (Binary/Timestamp/FixedSizeBinary/Utf8) instead of always Utf8, so the plan's output schema does not depend on which columns happen to be present. Regression test covers grouping by an entirely-absent FixedSizeBinary column (trace_id). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(server): rfc 0016 — skip the §7 default-limit injection for count-by queries apply_limit ran unconditionally, but compile::validate now rejects count+limit combined (RFC 0002 amendment 2026-07-15). Every aggregation query sent to the HTTP endpoint was therefore a clean 400. Skip the injection when a Stage::Count is present. Regression test confirmed via revert: fails without the fix, passes with it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * docs(querier): rfc 0002 — execute_aggregate doc names the tenant backstop scan input Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(querier): rfc 0002 — checked_add for the excluded-row tally Matches the existing pattern on rows: an overflow surfaces as an error rather than silently wrapping in release builds. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(querier): rfc 0002 — mark AggregateGroup non_exhaustive Matches QueryResult/QueryStats' convention for public response types. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(querier): rfc 0002 — reject i64-overflowing bucket widths at validate time bucket_expr's execution lowering casts the width to i64, but validate_group_terms only checked positivity — a width between i64::MAX and u64::MAX ns passed validation and failed later during planning with a different error path. Moved into validate() for one compile-time contract. Regression test confirmed via revert. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Run #4's L4 pair plateaued at 11,053/11,523 rows across every 10s poll instead of climbing to completeness. The picker's row ceiling (run #3's fix) had already ruled out "too large to finish in time" — the count never moved at all, which points at a cache serving the same stale answer on every retry rather than a slow ingest. Loki's bundled local-config.yaml enables the embedded results cache for query_range's metric/matrix path (L4's loki_query_matrix), keyed by the query+start+end+step tuple that loki_measure_frequency_pair repolls unchanged. The first (still-incomplete) response gets cached and echoed back on every subsequent poll. Plain log queries (loki_query_range, used by L1-L3/L6) aren't extent-cached the same way, so they self-heal across polls untouched by this. -query-range.cache-results=false trades Loki's own query latency for correctness of the harness's completeness poll — in Loki's favour, same as the other operator-tuning flags already on this container. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y

* feat(bench): rfc 0031 l4 — wire into the live dispatch loop Live-wires PairClass::L4 into rfc0031_indicative_comparative_run, the #[ignore]d container-based dispatch test. The previous slice proved the L4 machinery (ourios_aggregate_answer, parse_loki_matrix, pick_frequency_pair, compare_aggregations) only at the fixture level, against a hand-built Loki matrix response — this slice makes it real against a running Loki container and the actual corpus. L4 is picked and measured as its own step, kept OUT of the `Picks`/`specs: Vec<PairSpec>` pipeline the L1/L2/L3/L6 classes share: an aggregation's (bucket, group) -> count map is not a LineKey multiset, and forcing it through OuriosAnswer/compare_lines would misrepresent the state rather than model it (the same "make invalid states unrepresentable" reasoning the miner/parquet layers already follow). Concretely: pick_frequency_pair runs post-store-build like pick_template_pair; its PairSpec is built with the exact dsl/logql shape the fixture-level test already pinned; loki_query_matrix issues a real query_range metric call with `step` pinned to the bucket width so evaluation instants land on parse_loki_matrix's documented bucket-alignment convention (t = bucket_start + width); loki_measure_frequency_pair polls it to completeness the same way loki_measure_pair does for line-returning pairs. Both share the same Loki container and corpus replay as the existing pairs. Equivalence-required-but-bytes-unasserted: RFC0031.1 (result-set equivalence) is never optional, so run_l4_pair asserts compare_aggregations(...).is_equal() unconditionally — an L4 mismatch fails the run exactly like every other class's equivalence check. Only the bytes RATIO stays unasserted (M_L4 is still §7-DEFERRED, no frozen margin to gate against yet): print_l4_report reuses print_pair_bytes_gates, which already prints L4's ratio with no verdict. L4 is measured, equivalence-checked, and reported LAST — after the L1-L3/L6 evidence has printed and their frozen gates have already asserted — so an L4-only failure cannot destroy that evidence (the same run #11 salvage lesson the rest of the harness follows). A missing candidate is reported loudly at pick time, never silently skipped. Purely additive: class_pair_specs, build_pair_specs, frozen_gate_failures, print_pair_bytes_gates, print_indicative_report, PairSpec, and PairClass are unchanged — no frozen-gate behavior for L1/L2/L3/L6 is touched. Verification: cargo fmt --all --check, cargo clippy --all-targets --all-features -- -D warnings (workspace), cargo nextest run -p ourios-bench (165 passed, 7 skipped) and cargo test -p ourios-bench --all-features all green, including the untouched fixture-level rfc0031_5_l4_frequency_aggregation_bytes. The corpus-scale dispatch test itself needs Docker + OURIOS_COMPARATIVE_CORPUS, neither available in this sandbox — its first live proof is the comparative-bench dispatch workflow, same as every other slice in this harness's history. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — backtick-delimit the loki regexp argument The dispatch's first-ever run failed: capture_regex's own Go RE2 escapes (\s+, \S+) were embedded inside a double-quoted LogQL string literal, which tried to interpret those backslashes as its own escape sequences (\s is not a valid one) and Loki rejected the query with "invalid char escape" before the pattern reached the regex engine. Fixed by switching to a backtick-delimited (LogQL/Go raw string) regexp argument, which passes the pattern through literally. Extracted the duplicated PairSpec-construction block (present independently in the fixture test and the live-wiring loop) into one shared l4_pair_spec helper, closing the drift risk and centralizing the fix. Added a backtick guard: a capture_regex containing a backtick (regex_escape does not escape backticks) would prematurely close the raw string, so the candidate is now rejected loudly instead of emitting a malformed query. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — measure L4 before, not after, the L1-L6 failure asserts The second dispatch run failed on the pre-existing, documented L3 Loki-side flake (0 of 9 rows before timeout) — but the run never even attempted L4: the failures.is_empty() assert for L1-L6's own salvaged measurement failures sat textually BEFORE the L4 measurement/report code, so any earlier pair's failure aborted the test before L4 was ever reached. This inverted the design intent (an L4-only failure should not destroy L1-L6 evidence, not the other way around). Moved L4's measurement to run immediately after the report prints, before the gate/failures assertions. run_l4_pair now pushes a Loki-side measurement failure (flake) into the same failures vec the other classes salvage into, instead of panicking immediately — so a flaky L4 measurement no longer aborts before the L1-L6 evidence is captured, symmetric with the fix for the reverse direction. A genuine L4 equivalence MISMATCH still hard-panics immediately, unchanged: RFC0031.1 equivalence is never optional, matching L1-L6's own compare_lines assertion. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — cap the picker's row-count ceiling at 100K The third dispatch got past the control-flow fix and genuinely measured L4 — but the picked candidate (a service's dominant, near-catch-all template) summed to ~971K matching rows, and Loki returned only 811,775 of them before the 300s poll deadline (the same budget every other class's loki_measure_pair uses). L4_MIN_ROWS was a floor with no ceiling, so the picker had no reason to prefer a smaller, still-meaningful candidate. Added L4_MAX_ROWS=100_000 (comfortable margin at the observed ~2.7K rows/s Loki throughput) to frequency_shape_rejection, so the picker moves on to a candidate the poll can actually finish measuring. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — disable loki's query-range results cache Run #4's L4 pair plateaued at 11,053/11,523 rows across every 10s poll instead of climbing to completeness. The picker's row ceiling (run #3's fix) had already ruled out "too large to finish in time" — the count never moved at all, which points at a cache serving the same stale answer on every retry rather than a slow ingest. Loki's bundled local-config.yaml enables the embedded results cache for query_range's metric/matrix path (L4's loki_query_matrix), keyed by the query+start+end+step tuple that loki_measure_frequency_pair repolls unchanged. The first (still-incomplete) response gets cached and echoed back on every subsequent poll. Plain log queries (loki_query_range, used by L1-L3/L6) aren't extent-cached the same way, so they self-heal across polls untouched by this. -query-range.cache-results=false trades Loki's own query latency for correctness of the harness's completeness poll — in Loki's favour, same as the other operator-tuning flags already on this container. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — correct the results-cache disable flag name Run #5 never got past container startup: `-query-range.cache-results` doesn't exist ("flag provided but not defined"), so Loki's /ready check timed out on a container that failed to start at all. Checked the pinned v3.5.3 source directly instead of guessing again: queryrangebase.Config.CacheResults is registered under the `querier.` flag prefix in roundtrip.go, not `query-range.`. Correct flag is -querier.cache-results=false. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — widen the loki poll deadline to 900s Run #6 (with the corrected -querier.cache-results=false flag from the prior commit) proved the results-cache theory wrong: L1-L3/L6 all measured cleanly, but L4 still plateaued — 10752/11523 rows (93.3%), even slightly worse than run #4's 95.9% pre-fix, and the shortfall varies run to run rather than repeating a fixed cached answer. That points at genuine, variable completion time rather than a bug: L4's LogQL runs a `| regexp` capture over every candidate line before grouping and counting, a real per-line cost the other classes' plain stream/count queries never pay. Widened loki_measure_frequency_pair's deadline from 300s to 900s — well inside the CI job's unset (360 min default) timeout given the whole run has taken ~95-100 min so far. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — raise loki's max-entries-limit, revert deadline theory Runs #4/#6/#7 all converged L4 to ~93-96% of expected rows, independent of poll deadline (300s vs 900s made no measurable difference) — ruling out both a results-cache echo (already disabled in #fe5915a) and a "just needs more time" theory (the deadline widening from the prior commit). A stable, time-independent shortfall points at something being permanently excluded, not merely delayed. Pulled the frozen otel-demo-v8 corpus locally and checked every log line matching the L4 pair's needle ("Wrote producer snapshot at offset") against its capture regex directly: all 11,525 matches parse cleanly. The regex/content isn't the problem — some matching lines are never being scanned at all. That points at Loki's default -validation.max-entries-limit (5000): count_over_time with a |regexp stage has to scan every raw kafka log line in a query-frontend split before the line filter narrows it down, and kafka's per-split volume exceeds 5000 lines often enough to silently truncate the scan before every matching line is reached. Raised the limit well past the corpus's noisiest single template's volume (~971K rows). Reverted the 900s deadline back to 300s (matching loki_measure_pair) — the widened deadline never addressed the actual bottleneck, and keeping it would misattribute the fix in a way that'd mislead the next reader. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — widen poll deadline now the entries cap is gone Run #8 (max-entries-limit raised) moved L4 from a hard ~93% plateau to 97.1% (11192/11523) — real progress, and unlike runs #4/#6/#7 the remaining gap now plausibly behaves like genuine ingest settle time rather than a fixed ceiling, since the artificial cap that made the prior 300s vs 900s test inconclusive is gone. Widened the deadline to 600s to test that directly before assuming a third factor is at play. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — revert unhelpful deadline widening, add diagnostics Run #9 (600s) measured 96.5% (11123/11523), statistically the same as run #8's 97.1% at 300s — deadline widening does nothing here, so the remaining shortfall after the entries-limit fix is a second stable cap, not settle time. Reverted the deadline back to 300s to match loki_measure_pair rather than keep an unjustified change. Wired the existing dump_loki_diagnostics helper (already used by loki_measure_pair on a deadline miss) into loki_measure_frequency_pair too — it's built around spec.logql + stats parsing, which is query-shape-agnostic, so it works unmodified for the matrix path. If L4 still falls short, the next run's failure carries the raw Loki stats (chunk-fetch counts, any warnings) instead of another guess. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — epoch-align the loki query window to bucket boundaries Run #10's diagnostics (wired in the prior commit) confirmed the L4 query itself is well-formed and Loki answers it successfully — no error, no chunk-fetch shortfall visible in the sampled response. That, combined with runs #6-#10 all converging to a stable ~93-97% regardless of poll deadline (300s/600s/900s all statistically indistinguishable), rules out both a timing race and a malformed query. The real mismatch: Loki's query_range evaluates a step-grid starting exactly at `start` (start, start+step, ..., end), but Ourios's own bucket(width) semantics are epoch-aligned (floor(ts/width)*width) — `min_effective_time_unix_nano` (the corpus's raw earliest timestamp) has no reason to already be a bucket-boundary multiple. Unless (end - start) is an exact multiple of the bucket width, the step-grid leaves a fractional sliver at the tail of the range with no evaluated window covering it at all — real, ingested, settled data that's simply never queried, independent of poll duration. That's exactly the shape every run has shown. Snap `start` down and `end` up to the nearest bucket-width boundary in l4_pair_spec — costs nothing (no data exists outside [min, max] to inflate the count) and guarantees the step-grid fully covers the range. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, local (non-container) rfc0031_comparative unit tests — 39 passed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * diag(bench): rfc 0031 l4 — add an ingest-vs-query-side split probe Run #11 (bucket-aligned query window) measured 96.6% (11133/11523) — narrower than the pre-fix ~93% plateau, but still in the same stable band as runs #6-#10, all independent of poll deadline, entries-limit, and now bucket alignment. Six straight dispatches without closing the gap means continuing to guess at query-side LogQL/config knobs isn't warranted anymore. Added a decisive probe: on a deadline miss, loki_measure_frequency_pair now also runs a PLAIN line-filter count (no count_over_time, no regexp) for the same needle + window via the new loki_query_range_uncapped (limit sized to expected_rows, unlike the shared loki_query_range's fixed 5000 cap — which is below this pair's 11523 expected rows and would itself lie about the count). If that plain count also falls short by the same margin, the shortfall is ingest-side (Loki never stored those lines) and no further query tuning will fix it; if it's ~complete, the loss is specific to the aggregation path. Diagnostic-only change — no behavior change to the measurement itself, just evidence gathering on the existing failure path. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, local (non-container) rfc0031_comparative unit tests — 39 passed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — reject high-frequency candidates prone to loki dedup Run #12's decisive diagnostic confirmed the L4 shortfall is ingest-side: a plain unaggregated line-filter count for the same needle+window came back just as short (11160/11523) as every aggregation-path attempt. Loki's ingester silently drops a log entry that collides with another on (timestamp, body) within the same stream — a drop invisible to the OTLP push response's partial_success (push_otlp already asserts that field is clean on every push in every run so far). No query-side fix was ever going to close this gap; the picker was choosing a candidate Loki structurally can't ingest identically. kafka's template_id=16 ("Wrote producer snapshot") fires roughly every 15s. A local exploration against the real frozen corpus (offline, no Loki container — pick_frequency_pair only touches Ourios's own pipeline) found candidates at much lower frequency clear of the same floors: template_id=60 ("Periodic task") at ~144s average cadence, ~10x the failing candidate's margin. Added L4_MIN_AVG_INTERVAL_SECONDS (100s) to frequency_shape_rejection as a durable picker rule, not a one-off override — this protects any future re-run of the picker against landing on another collision-prone high-frequency template, not just this specific dispatch. Updated two pre-existing tests (pick_frequency_pair_finds_a_moderate_ cardinality_group, rfc0031_5_l4_frequency_aggregation_bytes) whose synthetic sub-3s timelines — convenient for test speed, not meant to model real timing risk — tripped the new floor; scaled their timestamps 1000x (preserving cardinality/row-count/needle assertions unchanged) so they represent a realistic, non-collision-prone example. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, local (non-container) rfc0031_comparative unit tests — 40 passed (39 prior + 1 new: frequency_shape_rejection_enforces_the_average_ interval_floor). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * diag(bench): rfc 0031 l4 — dump loki's own container logs on a deadline miss Run #13's lower-frequency candidate (template_id=60, ~144s average cadence) still fell short (1144/1197), and a corpus-side check ruled out the leading theory entirely: every one of the 1197 matching records has a UNIQUE timestamp AND a unique body (verified via jq against the frozen otel-demo-v8 corpus locally) — zero exact (timestamp, body) collisions possible. Loki's documented dedup rule cannot be the mechanism here, which means it likely wasn't the full story for the prior candidate either, even though lowering the frequency floor did measurably help (17.5% loss -> 4.4% loss). Also checked push_corpus_to_loki/push_otlp end to end for a harness- side drop: the batching loop appends every non-empty corpus line's resource_logs to `pending` before any flush, with a final flush after the read loop — no line is skippable, and push_otlp's retry resends the identical Bytes payload, so no bug found there either. Everything checkable from the client side (query responses, corpus content, our own push code) is now ruled out or confirmed clean. The next place to look is Loki itself: on a deadline miss, loki_measure_frequency_pair now also dumps the Loki container's own stderr, filtered to warn/error/drop/reject/rate-limit/stream-limit lines — the ingester logs these for exactly the mechanisms still on the table (rate limiting, out-of-order rejection, stream-limit drops), none of which are visible in a query response or push_otlp's already- clean partial_success check. Diagnostic-only — no change to measurement behavior. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, local (non-container) rfc0031_comparative unit tests — 40 passed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * diag(bench): rfc 0031 l4 — scrape loki's discarded-samples metrics Run #14's level=warn/level=error stderr scan came back with zero matches in 6618 total lines — whatever is causing the L4 shortfall (still 1147/1197 with the lower-frequency candidate), Loki doesn't consider it log-worthy. That rules out rate limiting, out-of-order rejection, and stream-limit drops as commonly logged at WARN. (The first attempt at the stderr filter was a naive "contains warn" substring match, which drowned in false positives from query text like `severity_text="WARN"` appearing inside level=info lines — fixed to match on the `level=` field precisely.) Loki's distributor increments loki_discarded_samples_total/ loki_discarded_bytes_total (labeled by reason) even for discards that don't warrant a log line — its own dedicated counter for exactly this question. Added dump_loki_discard_metrics, scraping /metrics on a deadline miss (extracted as its own function, alongside dump_loki_diagnostics, to keep loki_measure_frequency_pair under clippy's line-count lint). Diagnostic-only — no change to measurement behavior. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, local (non-container) rfc0031_comparative unit tests — 40 passed. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * feat(bench): rfc 0031 l4 — documented completeness margin (§7, 2026-07-17) Sixteen dispatches (runs #1-#16) exhausted every mechanism checkable from the harness's side without ever reaching exact L4 completeness. Runs #13-#16, specifically, ruled out: query-shape artifacts (a plain line-filter count matched the aggregation-path shortfall exactly), Loki's documented same-(timestamp,body) dedup (zero exact collisions found via direct corpus analysis), interleaving between a genuine mid-corpus kafka restart's two service-instance periods (cleanly sequential), a harness-side push/batching bug (read end to end, none found), anything Loki logs at WARN/ERROR (zero matches bar one harmless startup transient), and Loki's own discarded-samples Prometheus accounting (zero discards of any kind, any reason). This matches an open, unresolved upstream Loki issue (grafana/loki#10658 and related): wide-time-range queries silently missing a small, consistent percentage of lines, with no error, no discard signal, and no maintainer-identified root cause. It's a documented, external, currently-unfixable characteristic of the comparison partner, not an Ourios or harness defect. Adds L4_COMPLETENESS_MARGIN = 0.90 (real headroom over the observed 3.9-4.4% loss band) and compare_aggregations_within_margin — narrowly scoped: it still hard-fails, at any margin, on Loki reporting MORE than Ourios for any cell or a cell absent from Ourios's own answer, the two signals that would actually indicate a correctness bug. Only aggregate under-counting up to the margin is tolerated. compare_aggregations (exact) is untouched and still gates the RFC0031.5 fixture-level test's synthetic Loki answer. Wires the margin into both loki_measure_frequency_pair's poll-complete threshold (accept short-of-exact within margin instead of always retrying to a hard timeout) and run_l4_pair's equivalence assertion. RFC 0031 amended: RFC0031.1's L4 clause now states the margin explicitly, and §7's L4-query-shape entry (previously open) is closed with the full evidence trail and the margin decision. M_L4 (bytes-read) stays deferred — this unblocks a measurement, it doesn't freeze that margin. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, ourios-bench lib tests (32 passed, comparative module) + local rfc0031_comparative integration tests (40 passed), mdbook build. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — margin check is total-level, not per-cell Run #17 validated the completeness-margin design and immediately refined it: the poll-completion check passed cleanly (1153/1197, 96.3%), but the equivalence check then hard-failed on a single cell landing 1 row OVER Ourios's count (114 vs 113 for one bucket/value) while the aggregate total stayed a solid under-count — consistent with the same step-grid boundary imprecision already characterized (a record landing in an adjacent bucket), not fabrication. The original compare_aggregations_within_margin checked "Loki > Ourios" per cell, which was too strict for that kind of noise. Refined to check for phantom cells (a (bucket, group_key) Loki reports that Ourios's own answer doesn't contain at all) and Loki's TOTAL exceeding Ourios's total instead — this still catches the failure mode that would actually indicate a bug (wrong regex or wrong bucket math would produce cells Ourios never produced at all, or push the total over) while tolerating single-cell boundary noise on keys both systems agree exist. Added a test for the exact run #17 shape (a single cell over, total still under, must pass) alongside the existing phantom-cell and net-overcount tests (renamed from "overcount" now that the check is total-level). RFC 0031 §7's L4 entry updated with the refinement and its rationale. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, ourios-bench lib tests (33 passed, +1 new) + local rfc0031_comparative integration tests (40 passed), mdbook build. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — per-group_key margin, address PR #536 review PR #536's code review (14 fresh findings across Copilot + CodeRabbit's post-run-18 passes) surfaced one substantive correctness gap and several real documentation/robustness issues in the completeness- margin work. All verified against current code before fixing. Substantive fix — cross-key redistribution gap (CodeRabbit, Major): compare_aggregations_within_margin's grand-total-only check (from the run #17 fix) let Loki over-count one group_key while under-counting another by the same amount and still read as 100% complete: Ourios {A: 100, B: 100} vs Loki {A: 190, B: 10} sums to a "complete" 200/200 while hiding A being fabricated to compensate for B being nearly lost. Refined to aggregate ourios/loki BY group_key first (summing each key across every bucket it appears in), then apply the phantom/overcount/margin checks per-key. This still tolerates run #17's exact shape (a single bucket's +1 doesn't change a key's own total across its buckets) while rejecting the redistribution a pure grand-total check missed. Added regression tests for both shapes. Also populates real per-key examples in mismatch reports (Copilot: the old design returned examples: Vec::new() on both mismatch paths, despite the function accepting examples_cap and RFC0031.1 calling for example keys on a failed comparison). Documentation/robustness fixes (all verified against current code, none required a runtime-behavior change beyond the fix above): - Two stale comments still asserted Loki's same-(timestamp, body) ingester dedup as the shortfall's mechanism, contradicting the nearby docs that say this was directly disproven and the true mechanism is uncharacterized (Copilot, 6 threads pointing at 2 real sites: the frequency_shape_rejection rejection message and one test comment — the other 4 threads were already-accurate historical narrative, verified and left alone). - Missing `//` justification comment on one #[allow(cast_precision_loss)] (CodeRabbit). - L4 picker silently continues when no viable candidate exists (l4_spec.is_none()) — only an eprintln, no run failure, despite L4 being a must-win class (CodeRabbit). Now pushes into `failures`. - PR description inaccurately described L4's equivalence assertion (exact compare_aggregations, ordered after the frozen gates) — rewritten to match actual behavior (margin-based, before the frozen gates, matching the run #11 salvage design already documented inline). - RFC 0031 §7's L4 entry claimed the picker "prefers the lowest- frequency viable candidate" — the actual algorithm is first-fit in ascending (template_id, param) order, not an exhaustive ranking (CodeRabbit, tagged Heavy Lift). Reworded to describe actual behavior and the deliberate scope decision (a real ranking pass would cost a query per candidate against a corpus with tens of thousands of templates; not built given first-fit has now found a validated candidate three real dispatches running). - Double-backtick delimiters for a LogQL code span containing literal backticks (CodeRabbit, MD038). Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, ourios-bench lib tests (34 passed, +2 new) + local rfc0031_comparative integration tests (40 passed), mdbook build. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — absolute row tolerance, not pure percentage Run #19 (the verification dispatch for the round-1 review fixes) found a real edge case in the per-group_key percentage margin: a group_key with exactly 1 total Ourios row, where Loki captured 0 (0%). A pure ratio has no meaningful middle ground at n=1 — it's binary, 0% or 100% — yet losing one isolated occurrence is fully consistent with the already-characterized ~4-8% aggregate loss rate this whole margin exists to tolerate. Converted the per-key check from a ratio (loki/ourios >= margin) to an absolute row tolerance floored at 1: ceil(ourios_key_total * (1 - margin)).max(1). This tolerates a cardinality-1 key losing its only row while still catching a real shortfall on a large key (100 rows, tolerance 10, losing 20 still rejects) — the phantom-cell and per-key-overcount hard-fail checks are unaffected. Two new regression tests cover both ends. Verified: cargo fmt --all --check, workspace cargo clippy -D warnings, ourios-bench lib tests (36 passed, +2 new) + local rfc0031_comparative integration tests (40 passed), mdbook build. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — margin comparator precision, review triage compare_aggregations_within_margin's tolerance formula (ceil(o*(1-margin)) .max(1)) was itself miscalibrated for small-but-not-1 totals, per two independent Copilot review threads (o=2 at 90%: tolerance=1 permits 50% completeness, not 90%). Replace the subtract-then-round row tolerance with a direct, epsilon-guarded comparison — loki_total >= ourios_total * margin — which also sidesteps a second bug the naive floor() fix introduced: 1.0 - 0.9 isn't exactly 0.1 in f64, so floor(40.0 * (1.0 - 0.9)) truncated to 3 instead of 4, tightening the tolerance at exact 90%-boundary cases (caught by the existing margin_comparison_tolerates_undercount_within_margin test's svcB case). Extract phantom_cells and aggregate_by_group_key helpers to bring the function back under clippy's line limit, and add a # Panics section for the margin-validation assert. Fix several accumulated PR #536 review findings: run_l4_pair's doc comment claimed L4 runs after the L1-L3/L6 frozen gates assert (it actually runs before, printing first); three "ingest-vs-query" overclaims (the plain line-filter probe still calls query_range, so it can rule out "specific to the metric-aggregation path" but not prove ingest-side loss); L4_COMPLETENESS_MARGIN's own doc comment still described the superseded total-level design. Add a clarifying comment on l4_pair_spec's step-grid reasoning (the first evaluated instant decodes to an empty phantom bucket, not a lost real one) and fix the RFC's LogQL code span, which kept the backslash escaping needed for single backticks even after switching to a double-backtick delimiter that makes it unnecessary. Reconcile RFC0031.5's must-win predicate with M_L4 staying deferred — add a note that the predicate is the target contract, not currently gated. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — panic-safe diagnostic probe, more review triage loki_query_range_uncapped used expect()/assert!() internally, but it runs on the L4 deadline-miss diagnostic path inside the same runtime.block_on that gathers L1-L3/L6's evidence — a panic there (a real Loki error response, a malformed body) would unwind the whole async block and lose all of it, defeating the print-before-assert salvage design (Copilot). Converted to return Result<u64, String> instead of panicking, matching the already-panic-free sibling diagnostics (dump_loki_diagnostics et al.). Also: fix a test comment that said "one row under the ceiling" for a fixture that actually lands exactly at the ceiling; fix an unreachable! message's imprecise invariant claim (the real gating condition is l4_spec.is_some() implies l4_loki.is_some(), not "iff frequency is Some"); document loki_query_matrix's whole-second/bucket-alignment precondition and verify it against l4_pair_spec, its only caller; reorder loki_measure_frequency_pair's deadline-miss diagnostics to run only when the completeness margin is actually missed, not on every deadline-miss regardless of outcome; fix two fixture comments claiming "~300s average spacing" that don't match their own timestamps (actually ~580s) and a comment attributing the L4 shortfall to ingest-side dedup after that theory was directly disproven elsewhere in the same file. Verified the remaining ~40 accumulated review threads (mostly a recurring "shared 300s deadline" doc/code mismatch and the run_l4_pair ordering claim, duplicated across many review rounds) against current code: all already correct, superseded by earlier commits in this investigation. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — round-4 review triage on the margin comparator Validate margin at function entry rather than after the phantom/overcount checks, so an invalid margin always panics per the documented contract instead of potentially returning a data-shaped mismatch first (CodeRabbit). Fix the doc comment paragraph still describing the superseded floor-based tolerance (Copilot). Reword the L4-skip diagnostic and failure message to name both reasons l4_spec can be None (picker bounds vs a backtick in the capture regex) and to make clear the skip fails the dispatch rather than reading as benign (Copilot, two sites). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): rfc 0031 l4 — margin=1.0 strictness + panic-safe matrix poll Two Copilot findings on the previous commit, both verified genuine: The cardinality-1 exemption applied at any margin, so a caller passing margin = 1.0 (exact completeness) would still accept Loki returning 0 of 1 for an n=1 key — the exemption now only applies to a genuinely fractional margin, with a regression test covering both directions at 1.0. Bit-identical behavior at the harness's 0.90. loki_query_matrix still used expect/assert internally, so a transient transport error, 5xx, or torn body during the L4 poll — which runs LAST in the same async block holding every other pair's already-collected measurement — would panic and unwind all of it. Converted to Result<L4Measured, String>; the poll loop now retries an Err until its deadline exactly like an incomplete answer, then surfaces it as the pair's failure. Extracted the below-margin shortfall diagnostics into dump_l4_shortfall_diagnostics to stay under clippy's function-length limit. Neither change alters the measured semantics run #23 is currently confirming (the comparator formula is untouched; at margin 0.90 the exemption gating is unchanged). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * docs(bench): rfc 0031 l4 — document why the phantom check is cell-level Copilot's latest pass proposed weakening phantom detection from (bucket, group_key) cells to bare group_keys so a boundary-exact record shifting into an empty adjacent bucket can't read as phantom. Declined: a systematic bucket-decode error (every cell shifted one width — the run #11 bug class) leaves every per-key total intact, so the cell-level check is the only guard that catches it, while the false positive it risks requires a nanosecond-exact bucket-boundary timestamp that no real dispatch has ever produced. Documented the trade-off on phantom_cells instead of changing behavior. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

* feat(bench): rfc 0009 d1/d2 sustained-ingest soak harness Adds `ourios-bench soak`: an in-process soak that drives the real ingest path — OTLP export -> IngestPipeline -> group-commit WAL fsync (100 ms window) -> miner -> Parquet record sink on a local Store — at a paced target rate, and samples the compaction backlog while driving `run_sweep` manually. The core mechanism is a synthetic clock: compaction only acts on sealed partitions (hour end + grace), which is time-driven logic, so record timestamps advance on a compressed timeline (`time_compression` synthetic seconds per wall second; default 60 = one wall-minute of load per synthetic hour) and the same synthetic now feeds `run_sweep`/`plan_candidates`. The sealing logic runs unmodified — only the timestamps it compares are compressed — so the real seal -> sweep -> compact path is exercised without waiting real hours. D1 throughput and ack latency are measured in wall-clock time and are unaffected; ack latency is taken at the `IngestPipeline::ingest` boundary (commit wait + in-order miner hand-off), a conservative upper bound on the WAL-commit latency and exactly what an OTLP client sees. The report (JSON via --out + stdout summary) carries the achieved and per-core rates, ack-latency percentiles, the backlog timeseries, and D1/D2 verdicts with their exact bars (>= 100_000 lines/s/core with p99 <= 200 ms; backlog bounded, returning to zero and draining to a final zero). A workflow_dispatch-only soak-bench.yml runs it on the ci-runner (indicative, non-authoritative) with a verdict-table job summary and the JSON as artifact. Hazard #3 (WAL durability vs. latency): the harness changes no ingest code; it measures the existing batched-fsync path with the production 100 ms window through a shared-Wal Journal wrapper, on a multi-thread runtime so the coordinator's spawn_blocking fsync offload behaves as in the server. Hazard #4 (small files): the sink flush target is deliberately small so sustained ingest produces the multi-file partitions compaction exists to consolidate, and D2 asserts the sweep drains them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): soak review round — pacing, bounded latencies, wall split, single-list backlog PR #558 review fixes, all accepted findings: 1. Load loop checks the deadline AFTER awaiting the tick, so a tick landing past the deadline schedules no extra batch (the 5 s smoke now acks exactly duration/pace batches). Test: load_loop_schedules_no_batch_past_the_deadline. 2. Both tickers pace with MissedTickBehavior::Delay: the sampler runs on an aligned interval (interval_at) instead of sleep-then-work, so the period no longer stretches by each sample's blocking time, and the load ticker no longer bursts back-to-back catch-up batches after a stall (which broke the paced-load assumption and inflated in-flight depth). 3. Ack latencies are bounded: LatencyRecorder caps stored samples at 2^22 and decimates by two on overflow, keeping a systematic 1-in-2^k sample of the stream (percentiles stay valid at bounded memory); the report carries latency_samples_stored/_total. Test: latency_recorder_decimates_to_a_systematic_sample. 4. wall_secs split into load_wall_secs (the D1 rate denominator, measured at end of load drain) and total_wall_secs (covers the sampler join + drain sweep, so no sample timestamp exceeds it); JSON, stdout summary, and the workflow jq updated. Smoke test asserts the ordering invariants. 5. D2Verdict::returned_to_zero_after_max docstring now states plainly that the post-load drain sweep reaching zero is itself the return-to-zero evidence; logic unchanged, pinned by an extended d2_verdict test. 6. Backlog bytes come from ONE tenant-wide listing per sample matched against the candidates' partition prefixes, not one listing per candidate — bounded list ops per sample on non-local stores. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * fix(bench): soak polish — write params in place, prompt sampler shutdown Two more accepted PR #558 review findings: 1. The batch generator writes each param into the body via `write!` (std::fmt::Write) instead of allocating an intermediate String per record slot in the hot loop. 2. Sampler shutdown is prompt: the stop flag rides a tokio watch channel and races the tick in a `select!`, so the sampler exits as soon as stop is flagged instead of idling out up to one full `sample_every_secs` tick (which inflated total_wall_secs for nothing — the drain sample already exists). The aligned-cadence / no-overlap properties are unchanged: samples still run strictly between ticks, one at a time. tokio grows the `macros` feature for the `select!`. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…high-water (#578) (#581) * fix(ingester): rotation stamp waits out the sweep's in-flight publish (#578) The age sweep's off-lock write_ordered opens a window where acked records exist only in the sweep's memory and the WAL: drained out of the sink buffers, not yet durable. A rotation firing in that window quiesced only the encode pool, saw drained-looking sinks, and stamped wal_high_water over the in-flight records — a crash before the store PUT completed then lost them (recovery replays only above the mark). Fix (#578's own direction, sibling of the RFC 0035 §3.1 encode barrier): the record sink carries a shared in-flight publish count (refcount + condvar, mirroring the encode pool's quiesce shape, panic-safe settling via a drop guard). PublishCoordinator::drain_aged/drain_all acquire the guard before the take, under the miner lock; Drained holds it until the off-lock write settles (durable, requeued, or unwound). Every wal_high_water stamping path goes through flush_then_snapshot, which now quiesces publishes first — order: quiesce encodes -> quiesce publishes -> flush -> stamp, with the class-by-class coverage argument in the comment. The count lives in SharedParquetSink shared state so any coordinator built over the sink feeds the same barrier, structurally. Touches §3.4 (WAL-before-ack / acked-data durability, hazard #3): the change only narrows when a snapshot horizon may advance — never past a record that is not durably in the store or back in a flushable buffer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * test(ingester): #578 fault-injection race + SIGKILL crash arm for the sweep barrier Two arms, together closing the acked-data-loss argument: - rfc0035_2_rotation_stamp_waits_for_the_sweeps_in_flight_publish (ourios-server): runs the sweep's two halves by hand — atomic drain under the miner lock, then a held-back off-lock write_ordered (the slow-S3-PUT seam) — while a real WAL rotation fires through the production rotation_snapshot_hook. Asserts mid-window that no snapshot is stamped, then that the stamp lands at the rotation mark only after the publish settles. Mutation-checked: reverting the quiesce_publishes in flush_then_snapshot fails the mid-window assertion deterministically (rotation completes in ~100 ms against the 800 ms observation point). - rfc0035_2_crash_during_the_sweeps_in_flight_publish_replays_the_records (ourios-ingester): extends receiver_sink_crash_fixture with a `sweep` window — acked records drained OUT of the buffers, write_ordered never started — SIGKILLs the process mid-window, and asserts recovery replays both records into the store. Proves the window's on-disk state is replayable; the race arm proves no stamp can land during the window to foreclose that replay. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * docs(rfcs): rfc 0035 status note — sweep-window #578 fixed, not open The status note called #578 a known open hazard outside the RFC's scope; the publish half of the §3.1 barrier now closes it (in-flight publish guard on the drains + quiesce in flush_then_snapshot). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * style(ingester): rename drain guard bindings (clippy used_underscore_binding) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * docs(server): barrier comment — post-recovery exclusivity + pool-optional quiesce Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y * docs(ingester): quiesce_publishes doc — exclusivity by construction on the recovery path Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WQY9wfrfRggqSpMLH8Xj3Y --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

jensholdgaard requested a review from Copilot April 28, 2026 13:27

Copilot started reviewing on behalf of jensholdgaard April 28, 2026 13:27 View session

Copilot AI reviewed Apr 28, 2026

View reviewed changes

Comment thread crates/ourios-core/Cargo.toml Outdated

Update crates/ourios-core/Cargo.toml

a3f1cbd

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Signed-off-by: Jens Holdgaard Pedersen <jens@holdgaard.org>

jensholdgaard merged commit 1bd0e59 into main Apr 28, 2026
6 of 7 checks passed

jensholdgaard mentioned this pull request May 2, 2026

chore(miner): add ourios-miner crate with §5 acceptance stubs (specified → red) #7

Merged

5 tasks

jensholdgaard deleted the chore/ourios-core-skeleton branch May 2, 2026 18:54

jensholdgaard mentioned this pull request Jun 11, 2026

docs(benchmarks): authoritative baseline thesis-gate results; rfc 0007 validated #182

Merged

jensholdgaard mentioned this pull request Jun 17, 2026

feat(server): wire the RFC 0014 data write path live (rfc0013 green .6) #245

Merged

This was referenced Jun 17, 2026

test(ingester): rfc0014 green .5 — crash no-loss (real SIGKILL) #246

Merged

feat(querier): rfc0018.4 green — event_name DSL filter #267

Merged

jensholdgaard mentioned this pull request Jun 30, 2026

epic: OTAP / direct Arrow ingest (explore) — columnar Arrow on the wire #324

Open

This was referenced Jul 11, 2026

fix(bench): RFC 0031 — byte-capped Loki push batching (run #2 finding) #476

Merged

fix(bench): RFC 0031 — 1.5 MB flush cap (Loki inflates OTLP internally) #477

Merged

jensholdgaard mentioned this pull request Jul 11, 2026

fix(bench): RFC 0031 — raise Loki's internal gRPC cap (single-line inflation) #478

Merged

jensholdgaard mentioned this pull request Jul 20, 2026

feat(bench): RFC 0009 D1/D2 sustained-ingest soak harness #558

Merged

Navigation Menu

chore(workspace): add ourios-core skeleton - #3

chore(workspace): add ourios-core skeleton#3
jensholdgaard merged 2 commits into
mainfrom
chore/ourios-core-skeleton

jensholdgaard commented Apr 28, 2026 •

edited

Loading

Uh oh!

jensholdgaard commented Apr 28, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jensholdgaard commented Apr 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Invariants and hazards touched

Test plan

Uh oh!

jensholdgaard commented Apr 28, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jensholdgaard commented Apr 28, 2026 •

edited

Loading