Bug #90: mds: don't sync log on every clientreplay request - CephFS - Ceph

Actions

Copy link

Bug #90

open

mds: don't sync log on every clientreplay request

Added by Sage Weil almost 14 years ago. Updated about 1 month ago.

Status:

New

Priority:

High

Assignee:

Category:

Performance/Resource Usage

Target version:

% Done:

Source:

Development

Tags:

Backport:

mimic,luminous

Regression:

Severity:

3 - minor

Reviewed:

Affected Versions:

ceph-qa-suite:

Component(FS):

MDS

Labels (FS):

task(easy)

Pull request ID:

Crash signature (v1):

Crash signature (v2):

Actions

Copy link

Updated by Sage Weil over 13 years ago

Target version set to 12

Actions

Copy link

Updated by Sage Weil over 13 years ago

Estimated time set to 5:00 h
Source set to 5

Actions

Copy link

Updated by Sage Weil over 13 years ago

Translation missing: en.field_position deleted (~~303~~)
Translation missing: en.field_position set to 302

Actions

Copy link

Updated by Sage Weil almost 13 years ago

Target version changed from 12 to 19

Actions

Copy link

Updated by Sage Weil over 12 years ago

Target version deleted (19)

Actions

Copy link

Updated by Sage Weil over 12 years ago

Translation missing: en.field_position deleted (~~480~~)
Translation missing: en.field_position set to 823

Actions

Copy link

Updated by Sage Weil over 11 years ago

Project changed from Ceph to CephFS
Category deleted (1)

Actions

Copy link

Updated by Sage Weil over 11 years ago

Translation missing: en.field_position deleted (~~1276~~)
Translation missing: en.field_position set to 581

Actions

Copy link

Updated by Greg Farnum almost 8 years ago

Category set to 47
Priority changed from Low to Normal

This'll probably kill us in a cluster of any size...?

Actions

Copy link

#10

Updated by John Spray almost 8 years ago

Is this definitely still an issue? I'm not seeing a particular place in the code where we do something different with the mdlog while in clientreplay

Actions

Copy link

#11

Updated by Greg Farnum almost 8 years ago

It might be fixed; we should check. But once upon a time (I think maybe still now?) then when in clientreplay set the flag that requests an immediate journal flush.

Actions

Copy link

#12

Updated by Zheng Yan almost 8 years ago

The code that flush journal. https://github.com/ceph/ceph/blob/8ce337e3552004fc4853c0c94f33235da4caa5df/src/mds/Server.cc#L1060 ?

Actions

Copy link

#13

Updated by Greg Farnum almost 8 years ago

Yeah. So, I think there was a reason for this. We definitely need to flush out any clientreplay requests before exiting clientreplay. But any ops in clientreplay shouldn't be dependent (except within a single-client stream, where we don't care), so our data guarantees should be met just by flushing once before we exit clientreplay. Right?

Or is there some other interaction we need to worry about?

Actions

Copy link

#14