Enable periodic cleanup of work_dir directories in ballista executor #1783
Conversation
| .context("Could not connect to scheduler")?; | ||
|
|
||
| let scheduler_policy = opt.task_scheduling_policy; | ||
| let cleanup_ttl = opt.executor_cleanup_ttl; |
There was a problem hiding this comment.
I wonder if it might be possible to use NamedTempFile or some other struct from tempfile (already a dependency): https://crates.io/crates/tempfile
There are at least two benefits:
- The files are dropped immediately after they are no longer used so required intermediate diskspace is minimized
- They can't be accidentally dropped while still in use (which perhaps affects long running queries)
I did something similar in DataFusion here: #1680
There was a problem hiding this comment.
@alamb Thanks for your advice 😊!
IMHP, if one job has 3 stage, stage2 read stage1 input then delete the file, but stage2 task fail,
In ballista, scheduler will start a task to reload stage1 input. I think using NamedTempFile will cause some trouble and complexity.
we need keep the file for task-recovery and stage retry (like spark). So i decide if all the files under job_dir not modified in TTL we can safely delete it.
If i am not right, Please correct me 🙈
There was a problem hiding this comment.
I don't honestly know enough about Ballista and its executor to know.
What do you think @yahoNanJing and @liukun4515 ?
|
@Ted-Jiang this is a great step toward making Ballista production ready 👍 Thanks for making the change. |
|
cc @liukun4515 |
| name = "executor_cleanup_enable" | ||
| type = "bool" | ||
| doc = "Enable periodic cleanup of work_dir directories." | ||
| default = "true" |
There was a problem hiding this comment.
Can we set the default value to false and open it after a long time of testing?
@Ted-Jiang
If the default is false, this feature will not impact the current status. Users who want to use this feature can open it.
Do you agree? @alamb
There was a problem hiding this comment.
I think it would be a good idea to put this feature in turned off by default so it can be tested more thoroughly
|
|
||
| [[param]] | ||
| name = "executor_cleanup_ttl" | ||
| type = "i64" |
|
Once someone who focuses on Ballista is happy with this PR I am happy to merge it |
I think @Ted-Jiang will address the comments and please hold the status. |
fix info Co-authored-by: Kun Liu <liukun@apache.org>
fix info Co-authored-by: Kun Liu <liukun@apache.org>
fix Co-authored-by: Kun Liu <liukun@apache.org>
liukun4515
left a comment
liukun4515
left a comment
There was a problem hiding this comment.
Thanks @Ted-Jiang
LGTM
5 participants
Which issue does this PR close?
Closes #1780 .
Rationale for this change
Avoid work_dir quickly fill up disk space.
What changes are included in this PR?
like spark standalone mode, Executor periodic spawn a task to clean work_dir, if all the files in job_dir not modified in executor_cleanup_ttl seconds, it will be deleted.
https://spark.apache.org/docs/latest/spark-standalone.html
spark.worker.cleanup.enabled