This is based on Jeff's most excellent work to identify why
non-recursive listing under postgreskv was phenomenally slow. It turns
out PostgreSQL's query planner was actually using two sequential scans
of the pathdata table to do its job. It's unclear for how long that has
been happening, but obviously it won't scale any further.
The main change is propagating bucket association with pathnames through
the CTE so that the query planner lets itself use the pathdata index on
(bucket, fullpath) for the skipping-forward part.
Jeff also had some changes to the range ends to keep NULL from being
used- I believe with the intent of making sure the query planner was
able to use the pathdata index. My tests on postgres 9.6 and 11
indicate that those changes don't make any appreciable difference in
performance or query plan, so I'm going to leave them off for now to
avoid a careful audit of the semantic differences.
There is a test included here, which only serves to check that the new
version of the function is indeed active. To actually ensure that no
sequential scans are being used in the query plan anymore, our tests
would need to be run against a test db with lots of data already loaded
in it, and that isn't feasible for now.
Change-Id: Iffe9a1f411c54a2f742a4abb8f2df0c64fd662cb
* put TestCreateV0 back in StoreForTest
* avoid direct handles to V0 pieceinfo db
* type mismatch fix
* use storage.Blobs interface in store_test.go
..instead of filestore.Store. this will allow filestore.Store to become
unexported.
* unexport filestore.Store
rename it to blobStore. things should use the storage.Blobs interface
instead. changes in this commit are purely mechanical (made through the
"refactor" tool in Gocode followed by search/replace on the word "Store"
within the storage/filestore/ directory).
* kill filestore.StoreForTest
now that filestore.blobStore is unexported, there isn't a need for a
specialized wrapper type. this (not coincidentally) also makes it
possible for the WriterForFormatVersion() method on
storagenode/pieces.StoreForTest to work, without requiring everything to
wrap the store.blobs attribute in a filestore.StoreForTest, which was
impractical.
We don't use reverse listing in any of our code, outside of tests, and
it is only exposed through libuplink in the
lib/uplink.(*Project).ListBuckets() API. We also don't know of any users
who might have a need for reverse listing through ListBuckets().
Since one of our prospective pointerdb backends can not support
backwards iteration, and because of the above considerations, we are
going to remove the reverse listing feature.
Change-Id: I8d2a1f33d01ee70b79918d584b8c671f57eef2a0
drpc will call Close on any transport we pass to it, but some
transports (like tls.Conn) will attempt to notify the remote
side of things. we don't want to do that, so pass a new
interface that just closes the underlying socket.
Change-Id: I53344d2747de21b3146abe4f82b8394bb8948cb5
* uplink/storage/streams: Upload loop ops reorganization
Reorganize the operations of the loop run by streamsStore.upload method
for not doing unneeded computations on each iteration.
* uplink/storage/streams: Move out returns values declaration
Move out return values declarations for those which aren't strictly
needed due to defer statements nor documentation purpose.
Change signature of metainfo DeleteObject to get rid of an extra call to
kvmetainfo GetBucket method and eliminate one round trip to the
satellite when deleting objects.
grpc doesn't exit dials right away if the context dialer
returns an error. since that's the only spot where we were
enforcing dial timeouts, dials could just leak for an
unknown amount of time.
add a timeout above the grpc dial because that's the documented
way that grpc expected to be canceled.
Change-Id: Ic47ac61ce8a5f721510cc2c4584f63d43fe4f2d5
