Improving {Halt,Retry}Error semantics

[GiedriusS]

It's a continuation of a small discussion that we had on Slack a few weeks ago. Currently, we have HaltError and RetryError. HaltErrors happen when some critical issue occurs and we have to halt the whole process so that the operator could investigate what went wrong. RetryError happens when some operation fails that we can retry. Some relevant snippets of code:

  366             // The RetryError signals that we hit an retriable error (transient error, no connection).                                                                                                         
  367             // You should alert on this being triggered too frequently.

  354             // The HaltError type signals that we hit a critical bug and should block                                                                                                                          
  355             // for investigation. You should alert on this being halted.

However, the way it is right now is if a RetryError occurs then we retry the whole iteration i.e. compactMainFn(). That's awesome but all of the previously performed work disappears because once we start that function again, we wipe the workspace:

https://github.com/thanos-io/thanos/blob/master/pkg/compact/compact.go#L938-L941
https://github.com/thanos-io/thanos/blob/master/pkg/compact/compact.go#L464-L469

This means that all of the files that were previously successfully downloaded have been lost.

I propose extending this logic to check what is on the disk before trying to overwrite the files. My idea is:

• Extend meta.json to include file integrity checksums of all files in the directory besides meta.json itself and deletion-mark.json. These should only be calculated by writers only if some flag has been specified so that this wouldn't incur a performance cost on all writers;
• Make Thanos Compact consult those checksums if they are present before trying to download files again. If they match then we don't download the file. If they don't - we overwrite and download the file again.

Affected Thanos versions: <=0.13.0.

Thoughts?

[pracucci]

My gut feeling is that your proposed logic adds extra complexity to cover a case which shouldn't be the normal one. I wouldn't expect failures to happen so frequently and, if they do, probably fixing the root cause is better, but I'm open to discuss it further.

Extend meta.json to include file integrity checksums of all files in the directory besides meta.json itself and deletion-mark.json. These should only be calculated by writers only if some flag has been specified so that this wouldn't incur a performance cost on all writers;

Some object storages already export the checksum of objects when reading attributes (HEAD operation). Have you checked if it's available for any object store we support? If so, you wouldn't need to store any checksum in meta.json.

[GiedriusS]

My gut feeling is that your proposed logic adds extra complexity to cover a case which shouldn't be the normal one. I wouldn't expect failures to happen so frequently and, if they do, probably fixing the root cause is better, but I'm open to discuss it further.

I actually had to implement a rough version of #2785 because it was impossible for Thanos Compactor to finish at least one iteration due to transient errors. RetryError should help out in this case but it wipes everything right now so it's pretty useless. It only "helps" by not restarting the whole process when --wait is specified.

If the object storage code doesn't support (and it shouldn't) retries when some more serious error happens and if our retrying logic is kind of useless IMHO then where do you think such things should be handled?

RetryError seems like an elegant solution to me but it needs improvements as outlined in the OP.

Extend meta.json to include file integrity checksums of all files in the directory besides meta.json itself and deletion-mark.json. These should only be calculated by writers only if some flag has been specified so that this wouldn't incur a performance cost on all writers;

Some object storages already export the checksum of objects when reading attributes (HEAD operation). Have you checked if it's available for any object store we support? If so, you wouldn't need to store any checksum in meta.json.

That's true but I don't think we should add another method for this to our object storage interface if we can avoid it because that raises the bar for any other object storage that might not support such functionality.

[pracucci]

That's true but I don't think we should add another method for this to our object storage interface if we can avoid it because that raises the bar for any other object storage that might not support such functionality.

We have the attributes function which was designed for this. Assuming it's feasible to add the checksum (I haven't checked if it's implemented by all storages we do support nowadays) I agree that adding extra attributes there could be a blocker to support other storages in the future.

[stale]

Hello 👋 Looks like there was no activity on this issue for last 30 days.
Do you mind updating us on the status? Is this still reproducible or needed? If yes, just comment on this PR or push a commit. Thanks! 🤗
If there will be no activity for next week, this issue will be closed (we can always reopen an issue if we need!). Alternatively, use remind command if you wish to be reminded at some point in future.

[GiedriusS]

WIP. Have basic hashes support in my own branch, now need to add the needed functionality in Compactor.

[stale]

Hello 👋 Looks like there was no activity on this issue for last 30 days.
Do you mind updating us on the status? Is this still reproducible or needed? If yes, just comment on this PR or push a commit. Thanks! 🤗
If there will be no activity for next week, this issue will be closed (we can always reopen an issue if we need!). Alternatively, use remind command if you wish to be reminded at some point in future.

[GiedriusS]

Attempting to implement this in #3031.

[stale]

Hello 👋 Looks like there was no activity on this issue for the last two months.
Do you mind updating us on the status? Is this still reproducible or needed? If yes, just comment on this PR or push a commit. Thanks! 🤗
If there will be no activity in the next two weeks, this issue will be closed (we can always reopen an issue if we need!). Alternatively, use remind command if you wish to be reminded at some point in future.

[stale]

Hello 👋 Looks like there was no activity on this issue for the last two months.
Do you mind updating us on the status? Is this still reproducible or needed? If yes, just comment on this PR or push a commit. Thanks! 🤗
If there will be no activity in the next two weeks, this issue will be closed (we can always reopen an issue if we need!). Alternatively, use remind command if you wish to be reminded at some point in future.

[GiedriusS]

Trying to push #3031 over the line.

[stale]

Hello 👋 Looks like there was no activity on this issue for the last two months.
Do you mind updating us on the status? Is this still reproducible or needed? If yes, just comment on this PR or push a commit. Thanks! 🤗
If there will be no activity in the next two weeks, this issue will be closed (we can always reopen an issue if we need!). Alternatively, use remind command if you wish to be reminded at some point in future.

[GiedriusS]

#3031 merged so closing 🎉

[locvo986]

Vào Th 3, 2 thg 3, 2021 lúc 04:15 Giedrius Statkevičius < [email protected]> đã viết:

Closed #2784 <#2784>. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2784 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ASCHX7A33GCKJ64SZJKZVCLTBP7WVANCNFSM4OCROJSQ> .

-- locvo