testing out if this is possible direction for performance imporvements #4444
Conversation
pkg/backup/item_collector.go
Outdated
|
|
||
| for _, group := range r.discoveryHelper.Resources() { |
There was a problem hiding this comment.
it might be worth making this put groups onto a channel and making it buffered that way we can control how many threads we spin out.
There was a problem hiding this comment.
Yeah, that's a good idea. So you're thinking we'll improve the performance of the retrieve operation by multi-threading? Definitely need to have a config somewhere that throttles the max number of ops/sec.
There was a problem hiding this comment.
For sure, this was to point to a diff when I come back and make an enhancement for this.
I will try and see if we can thread out the item_backupper and only have the tar writer as the blocking bit. This would also need a config for the max number of threads.
I just want to see if this is possible and, does it gain anything by doing a quick POC before making the enhancement.
If you want me to close the issue to make things neater, let me know :)
|
I wanted to add some specifics about the performance improvement with this change. The biggest benefit comes when a plugin takes some amount of time. (in my testing, I added a custom plugin that took 5 seconds). With this change, a backup of 13000 resources took 2hrs, and without this change, it took >13hrs. If we are going to make Protected Entities or something similar for DataMovers as plugins, this could be something to consider. @dsu-igeek WDYT? |
|
Going to close this draft PR, and move from POC to an enhancement design and discussion topic |
Thank you for contributing to Velero!
Please add a summary of your change
Does your change fix a particular issue?
Fixes #(issue)
Please indicate you've done the following:
/kind changelog-not-requiredas a comment on this pull request.site/content/docs/main.