[tamird]

With dependency tracking enabled, finalized dependency vectors were stored in TargetHash and cloned into HashFileData.dep_edges. Building that map also walked every hash entry a second time.

Return each finalized vector with its label hash and move it into dep_edges during aggregation. Instance ordering and dependency deduplication stay at the existing aggregation boundary. The serialized hash and edge formats remain unchanged; the removed field belongs to the internal Rust API. 1

Codex for Tamir.

[tamird]

This PR keeps the two maps. The prototype stored dependencies in each target row and projected the hash and edge JSON outputs from one map. It reduced finalization time and peak RSS with tracking on the measured leaf workload, but increased RSS without tracking and had inconclusive timing there.

The comparison used 1,000,000 unconfigured SourceFile leaves with zero dependency edges, Bazel -c opt builds with a MiMalloc test allocator, and fresh processes in baseline–candidate–candidate–baseline order for each mode on arm64 macOS. The common workload timed finalization, the production buffered hash writer, and tracked edge encoding.

Without tracking, the paired changes in the sum of measured phases were +5.47% and −11.78% relative to this PR, so time was inconclusive. Mean peak RSS increased from 432.0 to 473.4 MiB (+41.3 MiB, +9.57%). TargetHash grew from 104 to 128 bytes in this build; RSS covers the whole test process, including fixture setup and buffers.

With tracking, the sum of finalization, hash serialization plus writing, and edge encoding fell from 637.6 to 422.3 ms on average (−33.76%), lower in both pairs. Mean peak RSS fell from 567.4 to 475.0 MiB (−92.4 MiB, −16.28%). Edge encoding itself did not improve: 22.0 to 23.2 ms on average.

Hash outputs were 157,000,001 bytes and tracked edge outputs were 28,000,001 bytes; sizes and SHA-256 digests matched in every applicable run. This synthetic workload does not establish performance for representative graphs, parsing, or the full workflow.

Codex for Tamir.

[tinder-maxwellelliott]

This PR keeps the two maps. The prototype stored dependencies in each target row and projected the hash and edge JSON outputs from one map. It reduced finalization time and peak RSS with tracking on the measured leaf workload, but increased RSS without tracking and had inconclusive timing there.

The comparison used 1,000,000 unconfigured SourceFile leaves with zero dependency edges, Bazel -c opt builds with a MiMalloc test allocator, and fresh processes in baseline–candidate–candidate–baseline order for each mode on arm64 macOS. The common workload timed finalization, the production buffered hash writer, and tracked edge encoding.

Without tracking, the paired changes in the sum of measured phases were +5.47% and −11.78% relative to this PR, so time was inconclusive. Mean peak RSS increased from 432.0 to 473.4 MiB (+41.3 MiB, +9.57%). TargetHash grew from 104 to 128 bytes in this build; RSS covers the whole test process, including fixture setup and buffers.

With tracking, the sum of finalization, hash serialization plus writing, and edge encoding fell from 637.6 to 422.3 ms on average (−33.76%), lower in both pairs. Mean peak RSS fell from 567.4 to 475.0 MiB (−92.4 MiB, −16.28%). Edge encoding itself did not improve: 22.0 to 23.2 ms on average.

Hash outputs were 157,000,001 bytes and tracked edge outputs were 28,000,001 bytes; sizes and SHA-256 digests matched in every applicable run. This synthetic workload does not establish performance for representative graphs, parsing, or the full workflow.

Codex for Tamir.

Thanks for the metrics, approved

[tamird]

Green at last

[tamird]

@tinder-maxwellelliott one failure

FAIL: //src:rust_tests (Exit 101) (see /home/runner/.cache/bazel/_bazel_runner/789be8cce26f55b14079335a5cb85410/execroot/_main/bazel-out/k8-fastbuild/testlogs/src/rust_tests/test.log)

where is that log?

[tinder-maxwellelliott]

@tinder-maxwellelliott one failure

FAIL: //src:rust_tests (Exit 101) (see /home/runner/.cache/bazel/_bazel_runner/789be8cce26f55b14079335a5cb85410/execroot/_main/bazel-out/k8-fastbuild/testlogs/src/rust_tests/test.log)

where is that log?

Adding extraction of those failures here #546

[tamird]

This is green again!