Consider the following operations ordering:
1. Inhume(with tombstone A) --> add tombstone mark for an object
2. --> new epoch arives
3. --> GCMark is added for a tombstone A, because it is unavailable
4. Put(A) --> return error, because the object already has a GCMark
It is possible, and I have successfully reproduced it with a test on the
shard level. However, the error is related to the specific
_ordering_ of operations with engine. And triggering race-conditions like
this is only possible on a shard level currently, so no tests are
written.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
There might be situation when context canceled earlier than traverser move to another part of the nodes.
To avoid this, need to wait for the result from concurrent put at each traverser iteration.
Signed-off-by: Anton Nikiforov <an.nikiforov@yadro.com>
All error counting and hangling logic is present on the engine level.
Currently, we pass engine metrics with shard ID metric to shard, then
export 3 methods to manipulate these metrics.
In this commits all methods are removed and error counter is tracked on
the engine level exlusively.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
- `reportShardErrorBackground()` no longer differs from
`reportShardError()`, reflect this in its name;
- reuse common pieces of code to make it simpler.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Target container ID is taken from tombstone: cmd/frostfs-node/object.go:507
Also object of type `TOMBSTONE` contains objectID, so tombstone and
tombstoned object must have the same containerID.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
* Remove `relay` field from put streamer as it's no longer used;
* Fix initialization of `Relay` object writer parameter.
Signed-off-by: Airat Arifullin <a.arifullin@yadro.com>
If local EC chunk found, but remote node is off, then `HEAD --raw` request
returns object not found.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
* `Patch` can't be applied for non-regular type object (tombstones,
locks etc.)
* Complex object parts can't be patched. So, if an object has EC/Split
header, it won't be patched.
Signed-off-by: Airat Arifullin <a.arifullin@yadro.com>
Created grpc connection should be established, so perform Healthcheck request
to check connection is ok.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
For EC chunks need to return EC parent object ID as
EC chunks don't have own attributes but inherit parent's.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
FSTree file counter used by writecache. As writecache has now only one
storage, so it is required to use real object size to get writecache
size more accurate than `count * max_object_size`.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
Add tests for `CreateViper` and `ReloadViper` to ensure that no extra
files, except *.yaml, *.yml, *.json, are loaded from config directory.
Signed-off-by: Aleksey Savchuk <a.savchuk@yadro.com>
* Tree and object service have the same log for checking APE. So,
this check should be moved to common package.
Signed-off-by: Airat Arifullin <a.arifullin@yadro.com>
Do not fail on same latest version to run compact on upgraded metabase.
Use NoSync on compact.
Log every batch on bucket delete stage.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
Since `frostfs-cli control shards rebuild` command was added,
there is no need for background rebuild now.
For failover tests used used value 1 to rebuild only schema change.
Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
If slice is altered in `for` loop, we cannot use range over its
length: it may cause panic if slice gets shorter.
Signed-off-by: Ekaterina Lebedeva <ekaterina.lebedeva@yadro.com>
* Split the logic of write target initialization to different packages;
* Refactor patch and put services: since both service initialize the target
themselves.
Signed-off-by: Airat Arifullin <a.arifullin@yadro.com>
Since Go 1.22 a "for" statement with a "range" clause is able
to iterate through integer values from zero to an upper limit.
gopatch script:
@@
var i, e expression
@@
-for i := 0; i <= e - 1; i++ {
+for i := range e {
...
}
@@
var i, e expression
@@
-for i := 0; i <= e; i++ {
+for i := range e + 1 {
...
}
@@
var i, e expression
@@
-for i := 0; i < e; i++ {
+for i := range e {
...
}
Signed-off-by: Ekaterina Lebedeva <ekaterina.lebedeva@yadro.com>
The `TestGCDropsObjectInhumedFromWritecache` test was flaky because a
running asynchronous rebuild operation prevented GC from deleting the
object. A test-only shard option `WithDisabledRebuild` has been added
to fix this.
Signed-off-by: Aleksey Savchuk <a.savchuk@yadro.com>
exportloopref is deprecated.
gopatch:
```
@@
var index, value identifier
var slice expression
@@
for index, value := range slice {
...
-value := value
...
}
@@
var index, value identifier
var slice expression
@@
for index, value := range slice {
...
-index := index
...
}
@@
var value identifier
var channel expression
@@
for value := range channel {
...
-value := value
...
}
```
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Benchmark results:
```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/pilorama
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
│ old │ new │
│ sec/op │ sec/op vs base │
ForestSortedIteration/bbolt,root-8 207.2µ ± 6% 173.6µ ± 6% -16.23% (p=0.000 n=10)
ForestSortedIteration/bbolt,leaf-8 3.910µ ± 5% 3.928µ ± 7% ~ (p=0.529 n=10)
geomean 28.46µ 26.11µ -8.27%
```
They are not representative, as the worst case is when we have multiple
items of different lengths. However, `FileName` is usually less than 100
in practice, so the asymptotics is the same.
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
Renamed parameters `min/max` to avoid conflicts with
predeclared identifiers.
Replaced background context with parent context without
cancellation in closer functions in frostfs-node.
Signed-off-by: Ekaterina Lebedeva <ekaterina.lebedeva@yadro.com>
When node put chunk into EC container, `policer` may remove it as redundant.
This chunk marked as removed. When parent object removed and `gc` start iterating over chunk,
node count removing chunk twice.
Signed-off-by: Anton Nikiforov <an.nikiforov@yadro.com>
* Also, resolve dependencies and conflicts for object service
by creating stub for `Patch` method.
Signed-off-by: Airat Arifullin <a.arifullin@yadro.com>
The synchronized service reload protocol added in systemd version 253
requires that the service provides a MONOTONIC_USEC field alongside the
RELOADING=1 notification message for synchronization purposes. The value
carried in this field must be the system CLOCK_MONOTONIC timestamp at
the time the notification message was generated as systemd compares it
to other CLOCK_MONOTONIC timestamps taken by pid1.
Signed-off-by: Ekaterina Lebedeva <ekaterina.lebedeva@yadro.com>
This reverts commit 327d364f34.
Reverted due to the problem with reload signal sent by systemd.
`frostfs-ir` service reconfigures correctly and service's
statuses are being reported to systemd. However, since we
replaced `go:linkname` & `nanotime()` with `time.Since()`,
systemd refuses to accept reload signal response from
`frostfs-ir`. To maintain correct behaviour it was decided to
revevrt systemd-related changes until a better solution is
found.
Signed-off-by: Ekaterina Lebedeva <ekaterina.lebedeva@yadro.com>
