ExportXMLWordPrintableJSON

    • Type: Task
    • Resolution: Unresolved
    • Priority: Major - P3
    • None
    • Affects Version/s: None
    • Component/s: None
    • Storage Execution
    • Storage Execution 2026-11-23
    • None
    • None
    • None
    • None
    • None
    • None
    • None

      Current index build tests lack coverage for builds

      • exceeding ~120 MB of index keys
      • with more than 3 indexes in one createIndexes callĀ 

      Some mechanisms are only reachable at volume and are currently only tested at toy scale: merge fan-in, sorter behavior with many spilled runs, and memory accounting across many concurrent indexes. Others such as yields, restart, rollback, initial sync, disk-space handling, have small-scale correctness coverage, but their behavior on large builds (natural yields under concurrent traffic, recovery time proportional to build size, real disk footprint from spills) have little or no coverage.

      Gaps identified from a cursory review:

      • Merge fan-in: no test reaches the iterator cap (iteratorsMaxNum); merge behavior is never exercised.
      • Memory accounting: no test asserts bounded memory across concurrent indexes in one build.
      • Cache and eviction: no index-build test constrains the WiredTiger cache.
      • Disk: ENOSPC is only simulated (simulateAvailableDiskSpace failpoint, before/after vote-commit); spill-file footprint and cleanup after abort are unmeasured.
      • Replication and recovery timing: no test observes standby lag, flow control, or secondary catch-up as a function of build size, and recovery time after failover or initial sync with a multi-GB build in progress is unmeasured.

      Possible directions:
      Most of the gaps above could be reached within the current suites by turning the existing knobs down rather than building new infrastructure: lower maxIndexBuildMemoryUsageMegabytes, build 10-64 indexes instead of 3, and set a small cacheSizeGB.

      • A multi-index spill-heavy build (10+ indexes, lowered limit) asserting that merges happen and complete, and that memory stays bounded.
      • The same shape under a constrained cache, asserting the build finishes and eviction progresses.
      • Resumable variants that fail over mid-scan with many spilled runs, covering the merge/resume interaction.
      • Real-disk coverage for ENOSPC and spill footprint, if the suite can afford it; otherwise the failpoint-based tests remain the ceiling.
      • The timing gaps (replication pressure, recovery time, initial sync at scale) are measurements rather than pass/fail asserts; they likely would need a benchmark workload.
      • Stalls should be diagnosable from metrics, not bare timeouts

            Assignee:
            Stanislav Lukianov
            Reporter:
            Shin Yee Tan
            Votes:
            0 Vote for this issue
            Watchers:
            3 Start watching this issue

              Created:
              Updated: