Reflection data not purged

1

Hello,

we have some problems of executor volumes getting full quickly and found out some strange behaviour regarding the retention of reflection data.
Our observations are the following:

  • There is an accumulation of reflection data in the folders /opt/dremio/data/pdfs/accelerator/ for reflections that are deactivated. By deactivated I mean that the reflection is OFF in the UI and there is no materialization_id for this reflection in the sys.materializations table. We can however track the materialisation_id used in the Dremio logs of old reflection runs.
  • Large volumes of old data can also be found in the folders /opt/dremio/cloudcache/c0/ . Timestamps are after the reflections run so I guess it is linked…

Questions:

  1. Why is old reflection data not purged when the reflection is set to OFF ?
  2. In the documentation it is mentionned that “the cache will evict unused data when required” (Dremio). When is Dremio considering it “required” ?
  3. Is there a parameter somewhere in Dremio or an operation to perform in order to avoid such retention of unused reflection data ?
2
  1. Why is old reflection data not purged when the reflection is set to OFF ? - I assume these reflections were deleted 4 hours back, we have seen this once, do we the server.log the day these reflections were dropped?
  2. In the documentation it is mentioned that “the cache will evict unused data when required” (Dremio). When is Dremio considering it “required” ? - This is only done if new data comes in and needs space
  3. Is there a parameter somewhere in Dremio or an operation to perform in order to avoid such retention of unused reflection data ? - The default is 4 hours but is one value for all reflections