Josh Rosen created SPARK-13695:
----------------------------------
Summary: Don't cache blocks at different serialization level when
reading back from disk
Key: SPARK-13695
URL: https://issues.apache.org/jira/browse/SPARK-13695
Project: Spark
Issue Type: Improvement
Components: Block Manager
Reporter: Josh Rosen
Assignee: Josh Rosen
When a cached block is spilled to disk and read back in serialized form (i.e.
as bytes), the current BlockManager implementation will attempt to re-insert
the serialized block into the MemoryStore even if the block's storage level
requests deserialized caching.
This behavior adds some complexity to the MemoryStore but I don't think it
offers many performance benefits and I'd like to remove it in order to simplify
a larger refactoring patch. Therefore, I propose to change the behavior such
that disk store reads will only cache bytes in the memory store for blocks with
serialized storage levels.
There are two places where we request serialized bytes from the BlockStore:
1. getLocalBytes(), which is only called when reading local copies of
TorrentBroadcast pieces. Broadcast pieces are always cached using a serialized
storage level, so this won't lead to a mismatch in serialization forms if
spilled bytes read from disk are cached as bytes in the memory store.
2. the non-shuffle-block branch in getBlockData(), which is only called by the
NettyBlockRpcServer when responding to requests to read remote blocks. Caching
the serialized bytes in memory will only benefit us if those cached bytes are
read before they're evicted and the likelihood of that happening seems low
since the frequency of remote reads of non-broadcast cached blocks seems very
low. Caching these bytes when they have a low probability of being read is bad
if it risks the eviction of blocks which are cached in their expected
serialized/deserialized forms, since those blocks seem more likely to be read
in local computation.
Therefore, I think this is a safe change.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
