Productionize Linux jemalloc heap profiling

[kevinyang372]

Description

Productionizes automatic jemalloc heap profiling on Linux so high-memory events upload a heap profile to Sentry, matching what macOS already does.

The key difference from macOS: the Linux profile is uploaded unsymbolized (raw pprof). It carries sample addresses + memory mappings + the GNU build-id, and is symbolized offline against the debug-info file (DIF) the release pipeline already uploads to Sentry — the same artifact used to symbolize panics, matched by build-id. This lets the shipped binary stay fully stripped (no symbol-table bloat) while still producing usable, symbolizable profiles.

What changed

• app/src/profiling.rs: on Linux, dump the gzipped pprof in-process via jemalloc_pprof (dump_jemalloc_pprof_bytes) — no external pprof binary, HTTP server, or fixed port (essential for the headless remote-server daemon). The dump is now raw / unsymbolized. The external-pprof path is kept for macOS, where in-process symbolization stays.
• app/Cargo.toml: build jemalloc_pprof without the symbolize feature, so dump_pprof() returns a raw profile instead of symbolizing in-process (in-process symbolization would have required keeping the symbol table in the shipped binary).
• Cargo.lock: bump jemalloc_pprof / pprof_util to 0.8.2. This is required: 0.8.2 writes a usable mapping range (memory_limit = u64::MAX), whereas 0.8.1 wrote memory_limit = 0, leaving pprof unable to bind sample addresses to the binary — so profiles would be unsymbolizable even with the correct DIF. (See this PR: Use max limit in mappings polarsignals/rust-jemalloc-pprof#31)
• script/linux/bundle: enable jemalloc_pprof,heap_usage_tracking for the dev and preview channels, and keep the normal strip behavior (--strip-all for non-dev builds). The shipped binary stays small; symbols live only in the uploaded DIF.

Symbolizing a Linux heap profile (offline, via the Sentry DIF)

The release pipeline already uploads each build's debug-info file to Sentry (script/sentry_upload_dif.sh), keyed by GNU build-id. To analyze a heap-profile.pb from an "Excessive memory usage detected" event:

# 0. Use the standalone pprof (the Go-bundled `go tool pprof` misreads the
#    build-id on large DIFs):
go install github.com/google/pprof@latest

# 1. Read the main binary's build-id from the raw profile's mappings:
BUILD_ID=$(pprof -raw heap-profile.pb \
  | sed -n '/^Mappings/,/^Locations/p' \
  | grep -E 'warp-(dev|preview)' | head -1 | awk '{print $NF}')

# 2. Download the matching DIF from the channel's Sentry project
#    (a read-only token is sufficient; dev -> warp-client-dev,
#     preview -> warp-client-preview):
ORG=warpdotdev; PROJECT=warp-client-dev
ID=$(curl -s "https://us.sentry.io/api/0/projects/$ORG/$PROJECT/files/dsyms/?query=${BUILD_ID:0:20}" \
       -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" | jq -r '.[0].id')
mkdir -p difs
curl -sL "https://us.sentry.io/api/0/projects/$ORG/$PROJECT/files/dsyms/?id=$ID" \
  -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" -o "difs/${BUILD_ID}.debug"

# 3. Symbolize + analyze (pprof matches the DIF to the profile by build-id):
PPROF_BINARY_PATH=./difs pprof -http=: heap-profile.pb   # flame graph / top / graph

The DIF carries .symtab + DWARF, so frames resolve to function names with file:line and inlined frames. (addr2line -e difs/$BUILD_ID.debug <addr> works as a lower-level alternative.)

Notes / scope

• No pprof binary is bundled on Linux (in-process raw dump).
• stable Linux builds are unchanged (no profiling features) — enabling for dev + preview first.

Linked Issue

N/A — infrastructure/observability follow-up.

Testing

• cargo check + cargo clippy -- -D warnings for the app lib targeting Linux with heap_usage_tracking (exercises the new raw-dump branch); macOS host path also checked (validates the pprof_binary_path cfg-gate) — passes.

• ./script/format — no changes beyond the edited files.

• End-to-end validated on a build sharing this exact code path

• I have manually tested my changes locally with ./script/run

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-NONE

---

🤖 Generated with Warp Agent Mode
Conversation: https://staging.warp.dev/conversation/e6ba6343-1bb8-4cce-a155-6f0b453bad40
Plan: https://staging.warp.dev/drive/notebook/359wXMjgAzXhtOZfnSWUlA

[oz-for-oss]

@kevinyang372

I'm starting a first review of this pull request.

You can view the conversation on Warp.

I completed the review and no human review was requested for this pull request.

Comment /oz-review on this pull request to retrigger a review (up to 3 times on the same pull request).

Powered by Oz

[oz-for-oss]

Overview

This PR enables Linux dev and preview bundles to upload symbolized jemalloc heap profiles by using the in-process jemalloc_pprof dump path and preserving the executable symbol table when heap usage tracking is enabled. Stable Linux builds remain outside the rollout.

Concerns

• No blocking correctness, security, or spec-drift concerns found in the annotated diff.

Verdict

Found: 0 critical, 0 important, 0 suggestions

Approve

Comment /oz-review on this pull request to retrigger a review (up to 3 times on the same pull request).

Powered by Oz

[kevinyang372]

• Productionize Linux jemalloc heap profiling #12265 👈 (View in Graphite)
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.