Improve high-concurrency client throughput

[quinnj]

Summary

• remember origins that negotiate HTTP/1.1 during automatic protocol selection
• skip per-chunk write-deadline refresh when no write-idle timeout is active
• increase the response copy buffer from 8 KiB to 64 KiB
• replay buffered request bodies with an independent cursor over the same immutable backing storage

Root cause

Same-region Azure Blob profiles exposed four independent costs at high concurrency.

1. An HTTP/1.1-only origin caused every automatic request to try and fail HTTP/2 negotiation. These attempts serialized on the client HTTP/2 lock and opened a new TLS connection before falling back.
2. Each 16 KiB TLS write refreshed the descriptor deadline even when the request had no moving write-idle timeout.
3. The 8 KiB response copy loop refreshed the configured read-idle deadline too often and increased timer-heap contention.
4. Each buffered upload replay copied the complete request payload before the first attempt.

The HTTP/1.1 origin cache applies only to protocol = :auto. An explicit protocol = :h2 still performs HTTP/2 negotiation. Existing HTTP/2 origins keep the normal HTTP/2 path.

Benchmark evidence

The benchmark used one Julia process, 64 Julia threads, Julia 1.13.0-rc1, and an Azure Standard_E64ds_v6 VM with same-region Azure Blob storage. Protocol selection remained automatic.

• caching the HTTP/1.1 negotiation result raised the 1 MiB download profile from 0.43 to 3.93 Gbps
• the 64 KiB response buffer raised the same profiled case to 17.08 Gbps
• a 256 KiB buffer was slower at 16.63 Gbps, so this change keeps 64 KiB
• removing the buffered-body copy raised the 64 MiB upload case from about 21.3 to 23.7 Gbps
• the final full-stack 16 GiB matrix reached 38.8 Gbps median download and 26.6 Gbps median upload

Profile and Profile.Allocs were used to identify and verify the hot paths.

Validation

• full HTTP test suite passes on Julia 1.13.0-rc1 with 8 threads
• trim-compile suite passes: 63 of 63
• new automatic fallback integration test passes
• new deadline and buffered replay regression tests pass

Co-authored by Codex

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.55%. Comparing base (34882ba) to head (af91469).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1347      +/-   ##
==========================================
+ Coverage   88.45%   88.55%   +0.09%     
==========================================
  Files          31       31              
  Lines       11930    11955      +25     
==========================================
+ Hits        10553    10587      +34     
+ Misses       1377     1368       -9     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.