Alexandre Viau created CASSANDRA-21632:
------------------------------------------
Summary: BTI partition index preload is unbounded and throughput
is capped by page-granularity access
Key: CASSANDRA-21632
URL: https://issues.apache.org/jira/browse/CASSANDRA-21632
Project: Apache Cassandra
Issue Type: Bug
Components: Local/Startup and Shutdown
Reporter: Alexandre Viau
When a table has bloom_filter_fp_chance = 1, PartitionIndex.load warms the
entire Partitions.db by seeking to every 4096-byte offset and reading a single
byte. The warm is unbounded. Nothing relates the amount read to the amount of
page cache the node can actually retain
On our cluster (5.4 TiB and 32.5B partitions per node, bloom_filter_fp_chance =
1, Partitions.db totalling 309.47 GiB across 798 files, or 81,125,849 pages to
touch), roughly 70 GiB of page cache is available to the process. The warm
reads 309 GiB to retain at most a fraction of it, and startup takes roughly 90
minutes, essentially all of it in SSTableReader.openAll.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
