perf(sourcemap-upload): skip ZIP-level DEFLATE when wire codec compresses (62% CPU, 5.9% wire)

[BYK]

Closes #847.

Summary

The artifact-bundle ZIP was DEFLATE-encoded inside, then each wire chunk was re-compressed by the upload codec. Compressing already-compressed bytes is wasted work — DEFLATE has already extracted the redundancy that zstd/gzip would have extracted on the wire.

When the server advertises a wire codec, switch the ZIP to STORED so the wire codec does the work. Keep DEFLATE when no wire codec is available (the chunk-upload.no-compression kill-switch) so the upload doesn't balloon to raw bytes.

Measurements (CLI binary upload, 14.3 MiB raw)

	ZIP build	Wire encode	Total CPU	Wire bytes
Before (DEFLATE + zstd)	349 ms	4 ms	352 ms	3,818,608
After (STORED + zstd)	48 ms	85 ms	134 ms	3,592,161
Delta			−218 ms (62%)	−226 KB (5.9%)

Re-measured locally on master at 3f326bed. Same shape as the numbers in #847; the magnitude is slightly lower than the issue's 75% (because Bun's deflate is faster than the Python zipfile baseline I used in the original investigation).

Compatibility

• Server: reads the archive via zipfile.ZipFile in src/sentry/models/artifactbundle.py, which transparently handles STORED + DEFLATE entries. No protocol change.
• Symbolicator: SourceBundle::test() checks the SYSB magic header, not entry compression. No change.
• Pre-zstd / kill-switch servers (gzip-only or no compression): pickUploadEncoding returns gzip or undefined. STORED is only used when a wire codec is available, so plain uploads keep DEFLATE and stay small.
• Direct callers of ZipWriter.create(): default is compression: "deflate" — no behavior change.

Tests

• Property-based round-trip tests parameterized over both compression modes via describe.each.
• Unit tests reading the local-file-header method byte (0=STORED, 8=DEFLATE) on both ZipWriter and buildArtifactBundle.
• Sanity test asserting STORED is larger than DEFLATE on redundant input — would fail if ZIP layout drifts or input compressibility regresses.
• Existing tests (27 in zip + sourcemaps test files; 56 across all sourcemap-related suites) continue to pass.

Implementation notes

• New ZipCompression type exported from src/lib/sourcemap/zip.ts.
• pickUploadEncoding is lifted from inside uploadArtifactBundle to the top of uploadSourcemaps so the ZIP construction can see the codec choice. The encoding is then passed through to both buildArtifactBundle (for STORED-vs-DEFLATE) and uploadArtifactBundle (for wire compression).
• No env var escape hatch — the conditional on pickUploadEncoding(serverOptions.compression) is itself the safety net for any deployment that opts out of wire compression.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://cli.sentry.dev/_preview/pr-849/
Built to branch gh-pages at 2026-04-25 19:40 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[github-actions]

Codecov Results 📊

✅ 6007 passed | Total: 6007 | Pass Rate: 100% | Execution Time: 0ms

📊 Comparison with Base Branch

Metric	Change
Total Tests	📈 +57
Passed Tests	📈 +57
Failed Tests	—
Skipped Tests	—

All tests are passing successfully.

❌ Patch coverage is 32.56%. Project has 12899 uncovered lines.
✅ Project coverage is 75.57%. Comparing base (base) to head (head).

Files with missing lines (2)

File	Patch %	Lines
src/lib/api/sourcemaps.ts	17.86%	⚠️ 23 Missing
src/lib/sourcemap/zip.ts	60.00%	⚠️ 6 Missing

Coverage diff

@@            Coverage Diff             @@
##          main       #PR       +/-##
==========================================
+ Coverage    74.84%    75.57%    +0.73%
==========================================
  Files          285       286        +1
  Lines        53047     52804      -243
  Branches         0         0         —
==========================================
+ Hits         39699     39905      +206
- Misses       13348     12899      -449
- Partials         0         0         —

---

Generated by Codecov Action