tanmayrauth commented on issue #1096:
URL: https://github.com/apache/iceberg-go/issues/1096#issuecomment-4656700687
Thanks @C-Loftus, the sequencing looks right. A few specifics so the first
PR has a clear target:
1. Yes, regarding condensing HadoopCatalogFS. Several methods are
POSIX-isms a blob store can't/needn't provide: Mkdir/MkdirAll are no-ops on
object stores, and ReadDir overlaps with the ListableIO.WalkDir we already
require. The natural minimal surface is List + Stat + Read + Write +
Remove(All) + Rename. Worth noting today's gocloud blob IO implements only 3 of
the 9 required capabilities, so shrinking the interface (without changing
local-FS behavior) is what makes blob backends wireable at all.
2. On arbitrary schemes, I'm on board as a follow-up, but a doc note won't
be enough. The load-bearing method is Rename, which can't be atomic on S3 — any
shim is copy+delete, non-atomic even for a single writer (a crash mid-commit
leaves a half state). I'd gate the unsafe path behind an explicit opt-in on
CreateCatalog (e.g. allow-unsafe-commits=true, erroring otherwise) rather than
prose in docs. That preserves the safe fail-closed default we get for free
today.
3. Agreed on deferring the per-backend conditional writes (S3 If-Match,
GCS generation preconditions) — that's the real fix and shouldn't block the
plumbing.
4. Happy to spin up a separate issue to track the concurrent-write work.
Worth referencing the Java HadoopCatalog's own object-store warning there so
the constraint is documented.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
