Issue #4549
closed
Removing docker manifests from a docker repository takes a long time
Description
Removing all docker manifests from a large docker repo seems to take a long time:
~300 manifests takes ~2 minutes
~2000 manifests takes ~30-40 minutes
Reproducer:
1. Create and sync a docker repo such as: https://quay.io datawire/ambassador
2. Remove all docker manifests from the repo: pulp-admin docker repo remove manifest --repo-id=1-docker-dev-7915f7d0-7a98-4131-9c41-1be7b578d442 --not id=foo
Files
Related issues
Updated by ipanova@redhat.com about 5 years ago
- Triaged changed from No to Yes
- Sprint set to Sprint 50
We need to investigate this.
Updated by dkliban@redhat.com about 5 years ago
- Status changed from NEW to ASSIGNED
- Assignee set to dkliban@redhat.com
Updated by dkliban@redhat.com about 5 years ago
- Status changed from ASSIGNED to NEW
- Assignee deleted (
dkliban@redhat.com)
Updated by daviddavis almost 5 years ago
- Sprint changed from Sprint 51 to Sprint 52
Updated by amacdona@redhat.com almost 5 years ago
- Status changed from NEW to ASSIGNED
- Assignee set to amacdona@redhat.com
Updated by amacdona@redhat.com almost 5 years ago
I was able to reproduce this with the busybox repository (I'll make a bigger vm to do this again with datawire/ambassador next week) and the slowdown is already apparent, taking about 2 minutes to remove all the manifest-lists.
I've attached the c_profile of that task. How to view: https://docs.pulpproject.org/dev-guide/debugging.html#analyzing-profiles
Potential workaround:¶
The queries for removing content are much more strenuous than adding content. A potential workaround could be to create a new repository and add only the manifest lists that should remain. If for some reason we can't work out a big improvement, I'll profile the workaround to compare.
Updated by dkliban@redhat.com almost 5 years ago
- Sprint changed from Sprint 55 to Sprint 56
Updated by amacdona@redhat.com almost 5 years ago
- Related to Issue #5161: Removing manifest_lists from a repository does not purge all newly unlinked manifests added
Updated by amacdona@redhat.com almost 5 years ago
There is a minor bug in current Pulp2 removal which will be fixed by this change.
https://pulp.plan.io/issues/5161
Because of that bug, this fix will include a very minor behavior change. More manifests and blobs will be recursively removed when removing manifest_lists that share manifests. The manifests and blobs that will now be removed would have been inaccessible by tags, so should not affect users.
Updated by bherring over 4 years ago
- Copied to Test #5181: Removing docker manifests from a docker repository takes a long time added
Updated by amacdona@redhat.com over 4 years ago
Time improvement¶
Removal of 210 Manifests (unpatched): 84 minutes
Removal of 210 Manifests (patched): 45 seconds
Removal of 2997 Manifests(patched): 1 minute 19 seconds
Removal of 4381 Manifests (patched): 1 minute 30 seconds
Memory Usage¶
This speed improvement comes as a tradeoff for memory usage. Unpatched, the memory usage of the celery process was fixed and did not increase with the size of the call. (This is because pulp worked through the task one content unit at a time.)
With the patch, memory usage will increase with the number of content units specified in the removal call.
Patched removal peak RES memory usage of celery process:
REMOVAL : PEAK RES (k)
30 manifests: 234444
145 manifests: 251908
210 manifests: 260616
2996 manifests: 792260
4381 manifests: 902488
(Note: these numbers can vary based on the specific content. Manifests that reference a higher number of blobs will use more memory, etc.)
Reference tasks: (high memory tasks with docker):
Sync datawire/ambassador (2926 tags): 317572
Resting celery: 75260
COPY 2926 tags: 443820
To reproduce:
Big repos from quay.io: datawire/ambassador, calico/typha
Start (or restart) 1 pulp worker between each task
Use `grep VmHWM /proc/$pid/status` to determine peak RES usage
Conclusion:¶
Because memory usage scales up with removal size (limited only by repo size), there is a theoretical memory problem for arbitrarily large docker repositories. However even using a very large repo (~50,000 content units) , the peak RES memory was below 1 GB. Therefore, my best guess is that this problem will not be a practical concern for real docker repositories in the wild.
Updated by amacdona@redhat.com over 4 years ago
- Status changed from ASSIGNED to POST
Updated by mmccune@redhat.com over 4 years ago
nice work on the optimization. IMO, the memory consumption for this call is a worthy tradeoff for the performance increases we are getting with your optimization. If you were quoting numbers in the 2-10G consumption I'd be concerned.
Added by amacdona@redhat.com over 4 years ago
Added by amacdona@redhat.com over 4 years ago
Revision 76f5894b | View on GitHub
Flatten queries for content unit removal
https://pulp.plan.io/issues/4549
Replace the recursive pattern with a fixed number of larger queries.
Additionally, reorder the content removal to "top down". This will fail more gracefully; failure leaves orphans (safe) rather than user-facing unlinked content (unsafe). This requires the additional plugin step of removing the explicitly given content, which is normally handled by pulp platform.
A side effect of this change is the correction of a bug that did not remove shared content, even if all linked content is removed. https://pulp.plan.io/issues/5161
Updated by amacdona@redhat.com over 4 years ago
- Status changed from POST to MODIFIED
Applied in changeset pulp_docker|76f5894b7c593eafc8498d5215cb7e517cd4624b.
Added by bherring over 4 years ago
Revision ecba1db4 | View on GitHub
Refactor of test_remove.py with changes from 4549
With the refactor of the docker importer's remove function to
increase performance, content removal needs to be functional verified.
The cases covered with content post-count verification for all units:
1. Remove all manifest_lists sequentially.
2. Remove all manifests sequentially.
3. Remove all blobs sequentially.
4. Remove all manifest_lists batch.
5. Remove all manifests batch.
6. Remove all blobs batch.
7. Remove some non-shared manifest lists.
8. Remove some non-shared manifest.
9. Remove some shared manifests lists and verify shared units are not
recursively removed.
10. Remove some shared manifests and verify shared units are not
recursively removed.
The fixture includes:
* 2 relatively independent manifest lists (no shared manifests,
no shared blobs between them)
* 2 manifest lists that share some (but not all) manifests, and those
manifest share some (but not all) blobs. This only requires the creation
of 1 manifest list that shares some content with one of the first
“independent manifest lists”.
* 2 relatively independent manifests
* 2 manifests that share (some but not all) blobs
In order to sync the content, each content unit must be recursively related
to at least 1 tag.
closes #5181
Updated by amacdona@redhat.com over 4 years ago
Applied in changeset 76f5894b7c593eafc8498d5215cb7e517cd4624b.
Added by amacdona@redhat.com over 4 years ago
Revision fbc90740 | View on GitHub
Flatten queries for content unit removal
https://pulp.plan.io/issues/4549
Replace the recursive pattern with a fixed number of larger queries.
Additionally, reorder the content removal to "top down". This will fail more gracefully; failure leaves orphans (safe) rather than user-facing unlinked content (unsafe). This requires the additional plugin step of removing the explicitly given content, which is normally handled by pulp platform.
A side effect of this change is the correction of a bug that did not remove shared content, even if all linked content is removed. https://pulp.plan.io/issues/5161
(cherry picked from commit 76f5894b7c593eafc8498d5215cb7e517cd4624b)
Added by amacdona@redhat.com over 4 years ago
Revision fbc90740 | View on GitHub
Flatten queries for content unit removal
https://pulp.plan.io/issues/4549
Replace the recursive pattern with a fixed number of larger queries.
Additionally, reorder the content removal to "top down". This will fail more gracefully; failure leaves orphans (safe) rather than user-facing unlinked content (unsafe). This requires the additional plugin step of removing the explicitly given content, which is normally handled by pulp platform.
A side effect of this change is the correction of a bug that did not remove shared content, even if all linked content is removed. https://pulp.plan.io/issues/5161
(cherry picked from commit 76f5894b7c593eafc8498d5215cb7e517cd4624b)
Updated by amacdona@redhat.com over 4 years ago
Applied in changeset pulp_docker|fbc9074078bf4c324c3820bd3201449410ba95ad.
Updated by dalley over 4 years ago
- Status changed from MODIFIED to CLOSED - CURRENTRELEASE
.
Flatten queries for content unit removal
https://pulp.plan.io/issues/4549
Replace the recursive pattern with a fixed number of larger queries.
Additionally, reorder the content removal to "top down". This will fail more gracefully; failure leaves orphans (safe) rather than user-facing unlinked content (unsafe). This requires the additional plugin step of removing the explicitly given content, which is normally handled by pulp platform.
A side effect of this change is the correction of a bug that did not remove shared content, even if all linked content is removed. https://pulp.plan.io/issues/5161
fixes #5161 fixes #4549