AI 3.8-15.1.6 Release Notes

Summary

A new cluster-info API endpoint provides a simpler way to query cluster status programmatically.

Cold-read performance is a major focus of this release. Cold-read behavior is improved by removing serialized S3 GET bottlenecks, reducing small-read fetch amplification, and improving worker-side read execution so read-heavy workloads spend less time stalled on the page store or self-loopback paths.
Read-path execution efficiency is further improved through Netty worker-pool tuning and related read-path optimizations, helping the system stay more stable under sustained read pressure.
S3 API semantic compatibility is materially improved. ListObjects behavior is brought closer to AWS S3 expectations, missing buckets now return NoSuchBucket instead of a generic server error, invalid range requests return the expected range error, and response headers such as Content-Encoding and Accept-Ranges are handled more correctly.
S3 read behavior in worker-owned and proxy scenarios is improved through zero-copy path simplification and better handling of worker-owned segments, reducing avoidable stalls in affected read flows.

S3 write cache reliability fixes address cache eviction races, multipart placeholder races, and pinned-empty-file recovery after restart.
A direct-memory leak in the S3 HTTP path is fixed, preventing large sustained-memory growth under load.
FUSE write hangs triggered by unrecoverable worker-side errors are fixed.
Hash-ring mismatch handling is corrected so ghost-worker states no longer cause duplicate segments or unnecessary UFS reads.
S3 API correctness fixes improve list semantics, missing-bucket handling, invalid-range handling, and response-header correctness.
Operational fixes improve Docker collect-info behavior and reduce noisy or misleading logs in FUSE deployments.

No breaking changes in this release.

No migration steps are required for this release.

Last updated 29 days ago