feat(web3): drop hugepages, fold the xmrig guard into monero - #4383
Closed
Tanguille wants to merge 10 commits into
Closed
feat(web3): drop hugepages, fold the xmrig guard into monero#4383Tanguille wants to merge 10 commits into
Tanguille wants to merge 10 commits into
Conversation
…ctoria-metrics-k8s-stack (0.89.0 ➔ 0.90.0) (#4380) Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com>
48 slots (16x concurrency ratio) was a guess; 14d telemetry shows slots hit 32 in 0.055% of samples and never higher. Capping at 32 frees 16 slots (~1.2G) back to the shared byte budget for KV cache headroom.
v2.2.0 adds review_verbosity, which patches the bundled prompt with a ~300-word target and moves minor/info out of prose into the findings array. The hand-written "keep review length proportional" rule in the system_prompt addendum now duplicates it, so drop it.
Upstream #426 landed in v2.1.11: system_prompt_file and system_prompt are concatenated (file first, inline second) instead of the file being silently dropped. The standing rules move to .github/pr-review-prompt.md, leaving the scalar to hold only the per-PR release-notes branch.
tool_evidence_memory is on by default and seeds a corpus section telling the model to reuse prior evidence instead of re-gathering. The addendum is appended after the bundled default, so "cite only sources fetched this run" overrode it and forced a re-fetch every incremental review. Narrow the rule to the failure it was written for: claiming "confirmed" for a source that was never read.
…lumbing 42 of the last 83 reviews stopped at exactly the default 4 tool calls and none exceeded it, so the call budget bound every second review while the 900s wall clock went unspent. All 20 run_command calls in that sample were git_diff_stat, git_diff_name_only or git_status_short, re-deriving a diff the corpus already carries; on #4370 that was half the budget, and the review then called an unterminated regex correct without ever opening .renovaterc.json5.
PR #4349 emitted "Release notes: no release notes apply" and "Tool Harness Findings: the tool loop issued no tool calls" on a two-file Ceph change. The release-notes instruction fired on every PR regardless of whether anything was bumped; gate it on a version change and tell the reviewer to drop empty sections.
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
 |
|
Overall Grade |
Security Reliability Complexity Hygiene |
Code Review Summary
| Analyzer | Status | Updated (UTC) | Details |
|---|---|---|---|
| JavaScript |  |
Aug 6, 2026 9:43p.m. | Review ↗ |
| Shell | |
Aug 6, 2026 9:43p.m. | Review ↗ |
Important
AI Review is run only on demand for your team. We're only showing results of static analysis review right now. To trigger AI Review, comment @deepsourcebot review on this thread.
vm.nr_hugepages withheld 2368Mi from every node's allocatable memory even at 0 replicas, while control-2/3 sat at 99% booked with a 150Mi exporter Pending. Drop the sysctl and move xmrig to regular pages. xmrig requests 2Gi against the ~2336Mi RandomX needs: post-reclaim headroom is 2389Mi/2373Mi, and the true figure parks the miner Pending forever behind preemptionPolicy: Never. /tmp gains a sizeLimit since the dataset now shares the container's memory limit rather than living in hugetlb. Move xmrig-guard to monero/guard and rename its resources. Metrics stay xmrig_guard_* : the KEDA triggers, five alerts, the dashboard and 180d of series all reference them. Tuning, replayed over 180d at the native 20s cadence: - recovery dwell 300->180s, +0.8pp/+1.1pp over 30d, no change above 65C - panic 68->67C, zero measured cost, caps the hottest safe sample at 66.8C - recover 62 / trip 65 / trip_dwell 60 held against a 612-candidate grid Nothing is loosened: real bursts peak at 69.8C against a 70C rating, and the rise is p99 1.35C/min, not the 1.1 the design was sized on. control-1's flapping was telemetry, not policy. cadvisor scrapes at 60s where the other six sources scrape at 20s and query_cpu dates the observation by min() across them, so it self-invalidated 465 times in 10.26d against 45 and 49 on the NVMe nodes, each latching a 600s recovery dwell. Its budget is now 180s and carried on the policy, so the freshness check, the anti-replay gap and the dwell reset cannot drift apart; the VM step follows or the samples would never be returned.
Tanguille
force-pushed
the
feat/web3-guard-hugepages
branch
from
August 6, 2026 21:43
15a3c13 to
f2e1c58
Compare
Owner
Author
|
Recreating against |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #4381. 4th layer: #4378 <- #4379 <- #4381 <- this.
Hugepages
vm.nr_hugepages: 1184withheld 2368Mi from every node's allocatable memory even at 0 replicas. xmrig has been 0/0 on all three nodes for 6d12h under the thermal guard;HugePages_Free == HugePages_Totaland cluster-widehugepages-2Mirequests were 0%.Unblocks
rook-ceph-exporter-control-2(150Mi, Pending 2h).Applies on next reboot. Nothing was applied to the cluster.
Coupled edits required by the removal:
--huge-pages->--no-huge-pages, bothhugepages-2Miblocks, the HugePages emptyDir, andIPC_LOCK(xmrig takes it only to raiseRLIMIT_MEMLOCKforMAP_HUGETLB).xmrig memory 32Mi/128Mi -> 2Gi/3Gi. Requested under the ~2336Mi RandomX needs: at the true figure the miner never schedules and
preemptionPolicy: Nevermeans it waits rather than preempts./tmpgainssizeLimit: 64Mibecause the dataset now shares the container limit instead of living in hugetlb.Net reclaim is 2368Mi idle, ~320Mi while mining. RandomX loses hashrate without hugepages; accepted.
Guard move
web3/xmrig-guard/app/->web3/monero/guard/, standalone ks.yaml folded intomonero/ks.yaml. Resources renamed toguard.Metrics stay
xmrig_guard_*. The KEDA triggers, 5 alert rules, the dashboard and 180d of series reference them.No
dependsOnadded: KEDA'sfallback: {failureThreshold: 1, replicas: 0}plus thecount == bool 1andtimestamp >= time() - 120gates mean an absent guard yields 0 replicas, so ordering cannot produce mining without a gate.Tuning
Replayed 180d of
node_hwmon_temp_celsiusat the native 20s cadence over a 612-candidate grid. Only 2026-07-27 14:02 onward reflects the deployed policy; before that the guard read all sensors rather thantemp1(~9C hotter), and 07-17..21 it emitted without enforcing.Nothing loosened. Measured bursts peak at p50 67.8 / p90 68.8 / max 69.8C against a 70C rating. Rise is p50 0.88 / p90 1.13 / p99 1.35 C/min, not the 1.1 the design was sized on; the budget test now uses 1.35 and pins the panic path, which governs on a fast ramp (68.69C at 75s) since the trip path models 68.04C above the panic limit.
Guard is working: zero samples >=70C while
safe=1. Before it existed, ungated mining put control-3 over its rating 51% of the time it mined, peaking at 86C.The prior
55.0%/38.8% vs 44.7%/31.5%figures do not reproduce on either sensor set. Direction holds, magnitudes were stale; restated in the doc.control-1 freshness
Flapping was telemetry, not policy: 465 query errors in 10.26d against 45 and 49 on the NVMe nodes,
source_age_secondsp99 102s against a 120s ceiling. cadvisor scrapes at 60s where the other six sources scrape at 20s andquery_cpudates the observation bymin()across them. Self-reinforcing at 1.7 errors/h idle vs 4.8/h with a miner, because the cadvisor join only engages when one is present. 85 of 118 mining bursts died under 5 minutes.Budget is now 180s for the CPU path, carried on the policy so the freshness check, anti-replay gap and dwell reset read one number and cannot drift. VM
stepfollows, or the samples would never be returned. Its thresholds replay at 93.8% safe against 75.8% observed.Follow-ups, not in this PR
evictionHard: 7%->1GiandsystemReserved: 1Gi->512Mi: another +1.54Gi/node.kubeReservedmust NOT be cut, measured/podruntimeis already above its 2Gi.nvme1Micron 7450 480GB runs ~10C hotter than the 980 PRO beside it and sets the gate on both nodes. Sun-hour idle baseline is 63-64C against a workload adding 8.5C in 7 min, so 68-72% of surplus goes unmined. Cooling it beats any threshold change.Verification
26/26 tests pass;
kustomize buildclean on web3, monero/guard, monero/xmrig, cnpg cluster.