[AksharP5]

Finalizing long render-farm exports briefly blocks the coordinator and allocates hundreds of MiB for a small MP4 header. Write each sample-size table into one sized buffer, preserving the existing header bytes and chunk offsets.

Closes #2950.

Export metadata	Header time, before → after	Peak process RSS, before → after
1 h, 30 fps	232.9 → 20.2 ms	180.8 → 59.1 MiB
2 h, 60 fps	706.1 → 52.2 ms	622.5 → 83.1 MiB
4 h, 60 fps	1408.6 → 108.6 ms	1134.2 → 139.4 MiB

Medians of five alternating fresh-process trials per case with production one-second media runs on Linux, i5-14400F, Bun 1.4.1. The two-hour case cuts header time by 92.6% and peak process RSS by 86.6%. RSS measures transient allocation pressure including the process runtime. The benchmark isolates header assembly; it does not measure total export speed. The linked issue contains the executable benchmark, ranges and baseline commit.

Validation: all 160 render-farm tests pass; sample-size assertions now cover both video and audio tables, asymmetric byte order and unsigned high bits. Scoped Biome and strict TypeScript checks pass. All benchmark headers match byte for byte. A complete four-second MP4 made from real H.264/AAC packets also matches byte for byte and decodes without errors: 240 video frames and 188 AAC frames. No live fleet or S3 deployment was run.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

This PR reduces temporary allocations when building MP4 headers.

• MP4 exports build sample-size tables with one buffer.

No actionable issues were found. Checks in this review were code-based; tests and benchmarks were not rerun.

_{Reviews (1) · Last reviewed commit: "perf(render-farm): cut peak memory durin..." · Reviewed by Greptile}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
cap-status	Skipped		Oct 22, 2026 7:54pm UTC