TrueCloudLab/frostfs-node
19
1
Fork 27
Code Issues 91 Pull requests 8 Projects Releases 1 Activity Actions

Allow to seal writecache after flush #886

Merged
fyrchik merged 1 commit from dstepanov-yadro/frostfs-node:feat/flush_and_disable_writecache into master 2024-09-04 19:51:05 +00:00
Conversation 0 Commits 1 Files changed 13 +309 -5
dstepanov-yadro commented 2023-12-22 06:55:43 +00:00
Member
Copy link

  1. Added new flag seal for command frostfs-cli control shards flush-cache ... --seal that changes writecache mode to read-only after flush.

  2. Fixed manual flush: now objects from DB will be deleted, previously only fstree object deleted.

  3. Fixed manual flush 2: now it exclusive locks modeMtx, so manual flush doesn't conflict with background flush

  4. Put and Delete methods now return error immediately, if writecache is changing mode, closing or flushing objects with manual invokation.

To check:

  1. Start k6 to write data

  2. Grafana shows that writecache stores objects

  3. Run frostfs-cli control shards flush-cache --id <shard_id> --endpoint ... --wallet ... --seal

  4. Grafana shows that writecache has stopped to store objects, all objects were flushed, mode was changed:
    image

UPD:

Also added frostfs-cli control shards writecache seal that does almost the same as frostfs-cli control shards flush-cache --seal, but moves writecache to degraded read only mode.

Output example:

$ frostfs-cli control shards writecache seal --all --endpoint ... --wallet ...
Enter password > 
Shard G9yBpG87G28VxvUdTY8e9M: failed with error "shard is in read-only mode"
Shard LhMy32HTvHAC8eSWArCwoh: OK
Total: 1 success, 1 failed
$ frostfs-cli control shards writecache seal --all --endpoint ... --wallet ....
Enter password > 
Shard NrZrnj9UreF2kZXX96AHLo: OK
Shard L6se643BehHdVpLZe444Hc: OK
Total: 2 success, 0 failed
$ frostfs-cli control shards writecache seal --id 6688MpNXDTdTMjgtcJEEtR --endpoint ... --wallet ...
Enter password > 
Shard 6688MpNXDTdTMjgtcJEEtR: OK
Total: 1 success, 0 failed

Relates #569

1. Added new flag `seal` for command `frostfs-cli control shards flush-cache ... --seal` that changes writecache mode to read-only after flush. 2. Fixed manual flush: now objects from DB will be deleted, previously only fstree object deleted. 3. Fixed manual flush 2: now it exclusive locks modeMtx, so manual flush doesn't conflict with background flush 4. `Put` and `Delete` methods now return error immediately, if writecache is changing mode, closing or flushing objects with manual invokation. To check: 1. Start k6 to write data 2. Grafana shows that writecache stores objects 3. Run `frostfs-cli control shards flush-cache --id <shard_id> --endpoint ... --wallet ... --seal` 4. Grafana shows that writecache has stopped to store objects, all objects were flushed, mode was changed: ![image](/attachments/f411a4ef-55a1-4888-b55a-ff76875c40a9) UPD: Also added `frostfs-cli control shards writecache seal` that does almost the same as `frostfs-cli control shards flush-cache --seal`, but moves writecache to degraded read only mode. Output example: ``` $ frostfs-cli control shards writecache seal --all --endpoint ... --wallet ... Enter password > Shard G9yBpG87G28VxvUdTY8e9M: failed with error "shard is in read-only mode" Shard LhMy32HTvHAC8eSWArCwoh: OK Total: 1 success, 1 failed $ frostfs-cli control shards writecache seal --all --endpoint ... --wallet .... Enter password > Shard NrZrnj9UreF2kZXX96AHLo: OK Shard L6se643BehHdVpLZe444Hc: OK Total: 2 success, 0 failed $ frostfs-cli control shards writecache seal --id 6688MpNXDTdTMjgtcJEEtR --endpoint ... --wallet ... Enter password > Shard 6688MpNXDTdTMjgtcJEEtR: OK Total: 1 success, 0 failed ``` Relates #569
image.png
60 KiB
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 33790b2d94 to 58702a57e7 2023-12-22 06:57:42 +00:00 Compare
dstepanov-yadro reviewed 2023-12-22 07:01:31 +00:00
pkg/local_object_storage/writecache/writecachebadger/flush.go Outdated
@ -239,3 +239,3 @@
// Write-cache must be in readonly mode to ensure correctness of an operation and
// to prevent interference with background flush workers.
func (c *cache) Flush(ctx context.Context, ignoreErrors bool) error {
func (c *cache) Flush(ctx context.Context, ignoreErrors, _ bool) error {
dstepanov-yadro commented 2023-12-22 07:01:31 +00:00
Author
Member
Copy link

Decided to drop badger implementation, so no changed here.

Decided to drop badger implementation, so no changed here.
fyrchik commented 2023-12-26 08:37:56 +00:00
Owner
Copy link

Don't understand -- if we dropped badger and rebased, why this change?

Don't understand -- if we dropped badger and rebased, why this change?
fyrchik commented 2023-12-26 08:44:59 +00:00
Owner
Copy link

Oh, I see now.

Oh, I see now.
fyrchik marked this conversation as resolved
dstepanov-yadro reviewed 2023-12-22 07:02:56 +00:00
pkg/local_object_storage/writecache/writecachebbolt/delete.go Outdated
@ -34,3 +35,3 @@
}()
c.modeMtx.RLock()
if !c.modeMtx.TryRLock() {
dstepanov-yadro commented 2023-12-22 07:02:56 +00:00
Author
Member
Copy link

Writecache is closing, changing mode or flushing. So no need to wait, this step will be skipped.

Writecache is closing, changing mode or flushing. So no need to wait, this step will be skipped.
dstepanov-yadro reviewed 2023-12-22 07:03:45 +00:00
pkg/local_object_storage/writecache/writecachebbolt/flush.go Outdated
@ -136,13 +139,11 @@ func (c *cache) flushSmallObjects(ctx context.Context) {
}
}
c.modeMtx.RUnlock()
dstepanov-yadro commented 2023-12-22 07:03:45 +00:00
Author
Member
Copy link

refactoring

refactoring
dstepanov-yadro reviewed 2023-12-22 07:05:46 +00:00
pkg/local_object_storage/writecache/writecachebbolt/flush.go Outdated
@ -293,2 +301,3 @@
return c.db.View(func(tx *bbolt.Tx) error {
for {
batch, err := c.readNextDBBatch(ignoreErrors)
dstepanov-yadro commented 2023-12-22 07:05:46 +00:00
Author
Member
Copy link

Batching added to not to store all objects in single slice for delete.

Batching added to not to store all objects in single slice for delete.
dstepanov-yadro reviewed 2023-12-22 07:06:37 +00:00
pkg/local_object_storage/writecache/writecachebbolt/put.go Outdated
@ -44,3 +44,3 @@
}()
c.modeMtx.RLock()
if !c.modeMtx.TryRLock() {
dstepanov-yadro commented 2023-12-22 07:06:37 +00:00
Author
Member
Copy link

Writecache is closing, changing mode or flushing. So no need to wait, this step will be skipped.

Writecache is closing, changing mode or flushing. So no need to wait, this step will be skipped.
fyrchik commented 2023-12-27 07:18:40 +00:00
Owner
Copy link

It seems like a separate change in behavior, it deserves a separate commit.

It seems like a separate change in behavior, it deserves a separate commit.
dstepanov-yadro commented 2023-12-27 08:42:19 +00:00
Author
Member
Copy link

Done

Done
dstepanov-yadro reviewed 2023-12-22 07:08:05 +00:00
pkg/local_object_storage/writecache/writecachebbolt/storage.go Outdated
@ -69,3 +69,3 @@
}
func (c *cache) deleteFromDB(key string) {
func (c *cache) deleteFromDB(key string, batched bool) {
dstepanov-yadro commented 2023-12-22 07:08:05 +00:00
Author
Member
Copy link

When flush started by external command, there is no need for batching, because deletion processes in single goroutine.

When flush started by external command, there is no need for batching, because deletion processes in single goroutine.
dstepanov-yadro requested review from storage-core-committers 2023-12-22 07:09:53 +00:00
dstepanov-yadro requested review from storage-core-developers 2023-12-22 07:09:55 +00:00
acid-ant requested review from anikeev-yadro 2023-12-22 07:17:05 +00:00
anikeev-yadro approved these changes 2023-12-22 07:28:33 +00:00
dstepanov-yadro changed title from Allow to seal writecache after flush to WIP: Allow to seal writecache after flush 2023-12-22 07:46:11 +00:00
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 58702a57e7 to 0605da30de 2023-12-22 08:14:02 +00:00 Compare
dstepanov-yadro reviewed 2023-12-22 08:17:01 +00:00
pkg/local_object_storage/writecache/writecachetest/flush.go Outdated
@ -68,4 +68,2 @@
require.Error(t, wc.SetMode(mode.Degraded))
// First move to read-only mode to close background workers.
require.NoError(t, wc.SetMode(mode.ReadOnly))
dstepanov-yadro commented 2023-12-22 08:17:01 +00:00
Author
Member
Copy link

It was used to stop background workers. Now background workers disabled with writecache option. Also setting read-only mode makes it impossible to delete objects from db.

It was used to stop background workers. Now background workers disabled with writecache option. Also setting read-only mode makes it impossible to delete objects from db.
dstepanov-yadro changed title from WIP: Allow to seal writecache after flush to Allow to seal writecache after flush 2023-12-22 08:36:05 +00:00
acid-ant approved these changes 2023-12-22 10:42:10 +00:00
acid-ant approved these changes 2023-12-22 10:50:00 +00:00
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 0605da30de to f49fefe3ad 2023-12-22 11:39:29 +00:00 Compare
elebedeva approved these changes 2023-12-25 15:19:38 +00:00
fyrchik reviewed 2023-12-26 08:40:24 +00:00
pkg/local_object_storage/writecache/delete.go Outdated
@ -33,3 +34,3 @@
}()
c.modeMtx.RLock()
if !c.modeMtx.TryRLock() {
fyrchik commented 2023-12-26 08:35:32 +00:00
Owner
Copy link

If writecache is doing something, why don't we just hang here? Anyway, if it is not related to sealing, how about doing it in a separate commit or before your changes -- it deserves a separate description.

If writecache is doing something, why don't we just hang here? Anyway, if it is not related to sealing, how about doing it in a separate commit or before your changes -- it deserves a separate description.
dstepanov-yadro commented 2023-12-26 09:58:48 +00:00
Author
Member
Copy link

Writecache is not mandatory, but it can take a lot of time to flush for example. Also if modeMtx is locked, then highly likely writecache will be unavailable for write. So put or delete can be processed with blobstor directly.

Writecache is not mandatory, but it can take a lot of time to flush for example. Also if modeMtx is locked, then highly likely writecache will be unavailable for write. So put or delete can be processed with blobstor directly.
pkg/local_object_storage/writecache/flush.go Outdated
@ -135,13 +138,11 @@ func (c *cache) flushSmallObjects(ctx context.Context) {
}
}
c.modeMtx.RUnlock()
fyrchik commented 2023-12-26 08:36:34 +00:00
Owner
Copy link

Pure refactoring, please, make it a separate commit.

Pure refactoring, please, make it a separate commit.
dstepanov-yadro commented 2023-12-27 05:23:17 +00:00
Author
Member
Copy link

Done.

Done.
fyrchik marked this conversation as resolved
pkg/local_object_storage/writecache/flush.go Outdated
@ -281,3 +281,2 @@
c.modeMtx.RLock()
defer c.modeMtx.RUnlock()
c.modeMtx.Lock() // exclusive lock to not to conflict with background flush
fyrchik commented 2023-12-26 08:40:14 +00:00
Owner
Copy link

What do you mean by "conflict with background flush"?
Both flushes do the same, here we just need to return after everything is flushed.

What do you mean by "conflict with background flush"? Both flushes do the same, here we just need to return _after_ everything is flushed.
dstepanov-yadro commented 2023-12-26 09:56:15 +00:00
Author
Member
Copy link

Background flush acquires RLock, so only one flush will be in process: background or manual

Background flush acquires RLock, so only one flush will be in process: background or manual
pkg/local_object_storage/writecache/storage.go Outdated
@ -77,1 +72,3 @@
})
var err error
if batched {
err = c.db.Batch(func(tx *bbolt.Tx) error {
fyrchik commented 2023-12-26 08:38:47 +00:00
Owner
Copy link

Again, an optimization, please, separate commit.

Again, an optimization, please, separate commit.
dstepanov-yadro commented 2023-12-26 09:55:05 +00:00
Author
Member
Copy link

No, it isn't an optimization: for manual flush we don't need batch.

No, it isn't an optimization: for manual flush we don't need batch.
fyrchik commented 2023-12-26 09:58:27 +00:00
Owner
Copy link

Why is manual flush single-threaded then?

Why is manual flush single-threaded then?
dstepanov-yadro commented 2023-12-26 10:00:25 +00:00
Author
Member
Copy link

To reduce complexity. No need to do manual flush as fast as possible.

To reduce complexity. No need to do manual flush as fast as possible.
fyrchik commented 2023-12-27 07:01:55 +00:00
Owner
Copy link

On the contrary 8 hours vs 1 hour can make a difference -- this is all done by a human.

On the contrary 8 hours vs 1 hour can make a difference -- this is all done by a human.
dstepanov-yadro commented 2023-12-27 08:37:23 +00:00
Author
Member
Copy link

Looks like that these values taken not from real life: flush works pretty fast now.

Looks like that these values taken not from real life: flush works pretty fast now.
fyrchik marked this conversation as resolved
fyrchik reviewed 2023-12-26 10:28:03 +00:00
pkg/local_object_storage/writecache/flush.go Outdated
@ -284,2 +283,3 @@
defer c.modeMtx.Unlock()
return c.flush(ctx, ignoreErrors)
if err := c.flush(ctx, ignoreErrors); err != nil {
fyrchik commented 2023-12-26 10:28:03 +00:00
Owner
Copy link

There are 2 cases to consider:

  1. DEGRADED shard mode.
  2. READONLY shard mode.

What is the behaviour in these situations? It looks like we should fail in both:

  1. In DEGRADED writecache is not opened and SSD can be missing.
  2. In READONLY blobstor is in READONLY too (likely for a reason, like dead HDD), so flushing won't be done.
There are 2 cases to consider: 1. DEGRADED shard mode. 2. READONLY _shard_ mode. What is the behaviour in these situations? It looks like we should fail in both: 1. In DEGRADED writecache is not opened and SSD can be missing. 2. In READONLY blobstor is in READONLY too (likely for a reason, like dead HDD), so flushing won't be done.
dstepanov-yadro commented 2023-12-27 05:31:50 +00:00
Author
Member
Copy link

And we will fail:

Line 27 in 8180a06
func (s *Shard) FlushWriteCache(ctx context.Context, p FlushWriteCachePrm) error {

And we will fail: https://git.frostfs.info/TrueCloudLab/frostfs-node/src/commit/8180a0664f05b5adedf399cbd8ad90f0f37115aa/pkg/local_object_storage/shard/writecache.go#L27
👍 1
fyrchik marked this conversation as resolved
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from f49fefe3ad to a1c438acd4 2023-12-27 05:22:13 +00:00 Compare
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from a1c438acd4 to 20e9670f9a 2023-12-27 08:41:42 +00:00 Compare
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 98fd3bb184 to 5a9f171804 2023-12-27 12:00:47 +00:00 Compare
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 5a9f171804 to 11de6264e0 2023-12-28 08:28:13 +00:00 Compare
fyrchik reviewed 2023-12-29 09:56:31 +00:00
cmd/frostfs-cli/modules/control/flush_cache.go Outdated
@ -16,0 +15,4 @@
Short: "Flush objects from the write-cache to the main storage",
Long: "Flush objects from the write-cache to the main storage",
Run: flushCache,
Deprecated: "Flushing objects from writecache to the main storage performs by writecache automatically. To flush and seal writecache use `frostfs-cli control shards writecache seal`.",
fyrchik commented 2023-12-29 09:51:55 +00:00
Owner
Copy link

s/performs/is performed by/

s/performs/is performed by/
dstepanov-yadro commented 2023-12-29 12:00:43 +00:00
Author
Member
Copy link

fixed

fixed
fyrchik marked this conversation as resolved
cmd/frostfs-cli/modules/control/writecache.go Outdated
@ -0,0 +53,4 @@
cmd.Printf("Shard %s: OK\n", base58.Encode(res.GetShard_ID()))
} else {
failed++
cmd.Printf("Shard %s: failed with error \"%s\"\n", base58.Encode(res.GetShard_ID()), res.GetError())
fyrchik commented 2023-12-29 09:53:22 +00:00
Owner
Copy link

\"%s\" -> %q?

`\"%s\"` -> `%q`?
dstepanov-yadro commented 2023-12-29 12:00:34 +00:00
Author
Member
Copy link

fixed

fixed
fyrchik marked this conversation as resolved
pkg/services/control/server/seal_writecache.go Outdated
@ -0,0 +29,4 @@
for _, r := range res.ShardResults {
if r.Success {
resp.Body.Results = append(resp.GetBody().GetResults(), &control.SealWriteCacheResponse_Body_Status{
Shard_ID: r.ShardID.Bytes(),
fyrchik commented 2023-12-29 09:55:51 +00:00
Owner
Copy link

Doesn't *r.ShardID work? We already have similar code in other parts.

Doesn't `*r.ShardID` work? We already have similar code in other parts.
dstepanov-yadro commented 2023-12-29 12:00:27 +00:00
Author
Member
Copy link

fixed

fixed
fyrchik marked this conversation as resolved
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 11de6264e0 to 5afad725d3 2023-12-29 12:00:24 +00:00 Compare
dstepanov-yadro force-pushed feat/flush_and_disable_writecache from 5afad725d3 to 581887148a 2023-12-29 13:06:02 +00:00 Compare
fyrchik approved these changes 2023-12-29 14:14:57 +00:00
fyrchik merged commit 581887148a into master 2023-12-29 14:33:13 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
anikeev-yadro
acid-ant
elebedeva
fyrchik
Labels
Clear labels   
P0

Highest

  
P1

High

  
P2

Medium

  
P3

Low

  
badger

Badger DB related issues

  
frostfs-adm

   
frostfs-cli

   
frostfs-ir

   
frostfs-lens

   
frostfs-node

   
good first issue

Good for newcomers

  
triage
  
Infrastructure

Infrastructure, Automation, CI/CD or DevOps related stuff

  
blocked

The issue or PR is blocked by something

  
bug

Something is not working

  
config

Configuration format update or breaking change

  
discussion

Talking about something in order to reach a decision or to exchange ideas

  
documentation

Documentation related

  
duplicate

This issue or pull request already exists

  
enhancement

New feature

  
go

Language features adoption

  
help wanted

Extra attention is needed

  
internal

   
invalid

Something is wrong

  
kludge

This is a kludge and we know it

  
observability

Related to metrics, tracing and logs

  
perfomance

Perfomance related

  
question

More information is needed

  
refactoring

Refactoring

  
wontfix

This won't be fixed

No labels
P0
P1
P2
P3
badger
frostfs-adm
frostfs-cli
frostfs-ir
frostfs-lens
frostfs-node
good first issue
triage
Infrastructure
blocked
bug
config
discussion
documentation
duplicate
enhancement
go
help wanted
internal
invalid
kludge
observability
perfomance
question
refactoring
wontfix
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
5 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: TrueCloudLab/frostfs-node#886
No description provided.
Delete branch "dstepanov-yadro/frostfs-node:feat/flush_and_disable_writecache"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?