Release 2.0.9

[Moonpuck]

New Features

Random Write Mode

Added disk-backed random write support on top of object storage, addressing performance and correctness issues for random writes, appends, O_TRUNC, and ftruncate workloads:

• How to enable: set --temp_dir=<local directory> to turn on random-write mode. Writes are first staged on the local disk and uploaded to OSS on flush.
• Full POSIX write semantics: arbitrary-offset random writes, O_APPEND (guaranteed to land at EOF), O_TRUNC, ftruncate, and mixed read/write workloads.
• Data integrity: CRC64 verification on the flush upload path; fsync waits for data to be uploaded before returning by default.
• Per-file size cap: new --random_write_max_file_size option (default 100 GiB; writes/truncates beyond the limit fail with EFBIG; configurable up to the OSS multipart capacity of about 48.8 TiB).
• Staging disk space management: staging disk space is released after each flush; the new --temp_dir_free_bytes option protects free space on the local disk.

Other New Features

• Log output to syslog is now supported via --log_dir=syslog.
• The bucket can be created automatically at mount time via the new --auto_create_bucket and --agentic_bucket options.

Performance Improvements

• Disk cache: optimized the cache engine with multiple background vCPU executors for the libaio engine. Throughput improves by up to 45% under 128-way concurrent reads of 128 KB small files, all served from the disk cache.
• Disk cache: with disk cache enabled, first-time sequential reads of large files in the psync scenario improve by about 40%.

Bug Fixes

• Disk cache: stale refills based on an outdated ETag are rejected after cache drop, with proper fallback on reopen failure.
• Added path refresh retry with ETag verification on the read path and background refill, preventing reads of stale objects.

Upgrade Notes

• Random-write mode is mutually exclusive with enable_appendable_object and the two cannot be enabled at the same time.
• Enabling random-write mode will significantly degrade sequential write performance on large files, because all writes are staged through the local disk and bounded by its throughput. Enable it only when random-write workloads (random writes, appends, O_TRUNC, or ftruncate) are actually required.
• Random-write mode relies on a local staging disk. Make sure the disk hosting --temp_dir has sufficient capacity, and tune the free-space protection options as needed.

[zhenjiaseu]

LGTM