Description
The Stage page in the UI is very slow when a large number of tasks exist (tens of thousands). The new work being done in SPARK-18085 makes that worse, since it adds potential disk access to the mix.
A lot of the slowness is because the code loads all the tasks in memory then sorts a really large list, and does a lot of calculations on all the data; both can be avoided with the new app state store by having smarter indices (so data is read from the store sorted in the desired order) and by keeping statistics about metrics pre-calculated (instead of re-doing that on every page access).
Then only the tasks on the current page (100 items by default) need to actually be loaded. This also saves a lot on memory usage, not just CPU time.
Attachments
Attachments
Issue Links
- causes
-
SPARK-28430 Some stage table rows render wrong number of columns if tasks are missing metrics
-
- Resolved
-
- links to
