[lcian]

Session.put now automatically uses resumable uploads under the hood for eligible uploads. An upload is eligible if:

• Its known remaining source size is at least resumable_threshold_bytes (default 256 MiB), measured before any client-side compression
• It's bytes or a seekable stream

Non-seekable streams fall-back to a regular, non-resumable, PUT.

When on-the-fly compression is needed, we spool the compressed payload to a temporary file before uploading; Client(spool_dir=...) customizes its directory. Otherwise, we use the buffer or stream directly.

Usecase gains the following kwarg, shown with its default:

Usecase(
    "debug_files",
    resumable_threshold_bytes=256 * 1024 * 1024,  # "none" disables resumables
)

Session.put also gains resumable_threshold_bytes for per-request overrides, defaulting to the value set on the Usecase.

Resumable uploads retry transient failures or writes that make no progress up to two times, with exponential backoff with base 2 seconds plus 1 second of jitter.
The retry count and backoff reset whenever a request makes progress.
This is combined with connection retries configured on the urllib3 pool, which are still effective in parallel.

Closes FS-389

[linear-code]

FS-389

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[jan-auer]

For a follow-up: This should receive compress_with and then match on the compression algorithm internally. Same for other places where we pass a compress bool.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c50f3fc. Configure here.}

[cursor]

Resumable retries skip connection failures

Medium Severity

is_transient does not treat exhausted connect retries or HTTP 500 as retryable, so the upload loop cancels the session instead of probing the offset and continuing. Pool connect retries use zero backoff, so a short network blip after progress can discard a large persisted prefix. 502/503/504 and read timeouts are retried, and a probe-before-resend makes retrying these failures safe.

Additional Locations (1)

• clients/python/src/objectstore_client/_resumable.py#L361-L367

 

^{Reviewed by Cursor Bugbot for commit c50f3fc. Configure here.}