Fix #5062: mt_quick_append_events returned {NULL} for an empty event array

[jeremydmiller]

Closes #5062.

The defect

array_length('{}', 1) is NULL in PostgreSQL, not 0. So

return_value := ARRAY[event_version + array_length(event_ids, 1)];

evaluated to NULL whenever mt_quick_append_events was called with no events, and the function returned a single-element array holding NULL. Npgsql then failed the read into long[]:

System.InvalidCastException: Cannot read a non-nullable collection of elements because the
returned array contains nulls. Call GetFieldValue with a nullable collection type instead.

Why it masked the real error

The cast failure surfaced from QuickAppendEventsOperationBase.PostprocessAsync — i.e. from inside OperationPage.ApplyCallbacksAsync, the loop that collects per-operation exceptions into the batch's exceptions list. A throw out of that loop lands in ExecuteBatchPagesAsync's catch, which rethrows the thrown exception and discards everything already collected. Callers therefore got an unrelated, non-retryable InvalidCastException instead of whatever actually made the append fail. On the reporter's system that dead-lettered Wolverine messages that would otherwise have been retried.

The fix

Three coordinated changes:

1. QuickAppendEventFunction — COALESCE(array_length(event_ids, 1), 0), so the empty case means what it says: zero events appended, final version = the stream's current version. This is the reporter's suggested fix and the actual correctness fix.

2. QuickAppendEventsOperationBase.PostprocessAsync — skip the field-0 read when the stream carries no events. Nothing needs post-processing there anyway, and this matters for databases whose function has not been migrated yet (AutoCreate.None + explicit migrations): they stop throwing the cast exception on the old function. The row is still consumed, so the reader stays aligned for the rest of the page.

3. ProjectionUpdateBatch.WaitForCompletion — the one caller in Marten that could reach the function with empty arrays (an Append side effect that ended up with no events; every other path already filters on Events.Any()). A StartStream still gets its mt_streams row.

Tests

TenantPartitionedEventsTests/Regressions/Bug_5062_empty_quick_append.cs, on the reported configuration (Conjoined + UseTenantPartitionedEvents + Quick):

• function_returns_a_non_null_version_for_an_empty_event_array — the issue's SQL repro. The function's parameter list varies with configuration (metadata columns, server timestamps, tag tables, bigint events), so the test reads the deployed signature back out of pg_get_function_arguments and feeds every array parameter an empty array. That pins the function itself rather than one config's hand-written call.
• empty_quick_append_operation_is_a_clean_no_op — the same empty append driven through Marten's own generated call site and result read, which is where the InvalidCastException actually landed.

Both fail on master with the exact reported InvalidCastException; both pass with the fix. Full TenantPartitionedEventsTests suite green on net9.0 (238 passed, 2 pre-existing skips).

Upgrade note

The function body changed, so existing databases will show one mt_quick_append_events update in the next schema migration. It is idempotent afterwards.

🤖 Generated with Claude Code