A user on the forum was seeing the error "bad message", which was not
very helpful. This case from the ext4 filesystem using the code EBADMSG
to indicate it detected an invalid CRC, suggesting disk corruption.
This change adds some explanatory information about probable disk
corruption to all errors coming from the (*blobInfo).Stat() call, which
is where storagenode fs corruption problems will usually manifest.
Refs: https://github.com/storj/storj/issues/5375
Change-Id: I87f4a800236050415c4191ef1a0fc952f9def315
Calling stat() (really, lstat()) on every file during a directory walk
is the step that takes up the most time. Furthermore, not all directory
walk uses _need_ to have a stat done on every file. Therefore, in this
commit we avoid doing the stat at the lowest level of
walkNamespaceInPath. The stat will still be done when it is requested,
with the Stat() method on the blobInfo object.
The major upside of this is that we can avoid the stat call on most
files during a Retain operation. This should speed up garbage collection
considerably.
The major downside is that walkNamespaceInPath will no longer
automatically skip over directories that are named like blob files, or
blob files which are deleted between readdir() and stat(). Callers to
walkNamespaceInPath and its variants (WalkNamespace,
WalkSatellitePieces, etc) are now expected to handle these cases
individually.
Thanks to forum member Toyoo for the insight that this would speed up
garbage collection.
Refs: https://github.com/storj/storj/issues/5454
Change-Id: I72930573d58928fa25057ed89cd4ec474b884199
Implement a new service to read retain filter from a bucket and
send them out to storagenodes.
This allows the retain filters to be generated by a separate command on
a backup of the database.
Paralellism (setting ConcurrentSends) and end-to-end garbage collection
tests will be restored in a subsequent commit.
Solves https://github.com/storj/team-metainfo/issues/121
Change-Id: Iaf8a33fbf6987676cc3cf74a18a8078916fe673d
Currently uploads can cause a lot of IOPS, reduce this by introducing a
in-memory buffer on-top of the file.
Change-Id: I5f4e3e01c0a36258271d180b922107de447bcb59
this commit updates our monkit dependency to the v3 version where
it outputs in an influx style. this makes discovery much easier
as many tools are built to look at it this way.
graphite and rothko will suffer some due to no longer being a tree
based on dots. hopefully time will exist to update rothko to
index based on the new metric format.
it adds an influx output for the statreceiver so that we can
write to influxdb v1 or v2 directly.
Change-Id: Iae9f9494a6d29cfbd1f932a5e71a891b490415ff
* 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.
* nicer flags
* fix concurrency
* add concurrent workers
* initialize things
* fix tests
* close retain service
* ensure we don't have workers working on the same satellite
* ensure things compile
* fix other compilation issues:
* concurrency changes
ran this with `go test -count=1000` and it passed all of them.
- we add a closed channel so that we can select on it with
context cancellation.
- we put a once in so we only close the channel once.
- every time the queue/running state changes, we have to broadcast
because we may want to wake up N pending Wait calls or other
concurrent workers.
- because we broadcast, we don't need to do the polling in Wait
anymore.
- ensure Run doesn't start multiple times so that we don't have
to worry about concurrent Close with multiple Runs.
- hold the lock while we start workers so that a concurrent Close
with Run can't decide that there's nothing started and exit
and then have Run start things.
- make sure to poll the closed/context channels through loops
or at the start of Run calls in case Close happens first.
- these polls should be under a mutex because they have a default
case which makes it possible to schedule such that Close hasn't
executed the channel close so it starts more work.
- cancel a local Run context when it's going to exit to make sure
that any retainPieces calls have a canceled context.
- hopefully enough comments to both check my work and help readers
digest what's going on.
Change-Id: Ida0e226a7e01e8ae64fa2c59dd5a84b04bccfbd7
* use the retain error class
Change-Id: I1511eaef135f98afd57b878e997e4c8a0d11cafc
* concurrency fixes again
- forgot to update the gc test to use the old Wait api.
- we need to drop the lock while we wait for the workers
to exit, because they may be blocked on the condition
variable
- additionally, we need to broadcast when we close the
signal channel because the state changed: they want
to wake up and exit.
Change-Id: I4204699792275260cd912f29aa73720f7d9b14b5
* undo my misguided rename
Change-Id: I6baffe1eb0434e260212c485bbcc01bed3250881
* remove pollInterval
* format paragraph more nicely
* move skew calculation into retain pieces
Add retain service on storagenode. This service runs retain jobs that have been queued by the storagenodes. Rather than running retain jobs during the grpc Retain() call, the grpc call queues a retain job to the retain service and returns immediately afterwards, removing a significant bottleneck in garbage collection.
