cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Juho Mäkinen <>
Subject Questions on LCS behaviour after big BulkLoad cluster bootstrap
Date Wed, 06 Jul 2016 06:56:01 GMT
Hello. I'm in the process of migrating my old 60 node cluster into a new 72
node cluster running 2.2.6. I fired BulkLoader on the old cluster to stream
all data from every node in the old cluster to my new cluster, and I'm now
watching as my new cluster is doing compactions. What I like is to
understand the LeveledCompactionStrategy behaviour in more detail.

I'm taking one node as an example, but all other nodes have quite same

There are 53 live SSTables in a big table. This can be seen both by looking
la-*Data.db files and also with nodetool cfstats: "SSTables in each level:
[31/4, 10, 12, 0, 0, 0, 0, 0, 0]"

If I look on the SSTable files in the disk I see some huge SSTables, like a
37 GiB, 57 GiB, 74 GiB, which are all on Level 0 (used sstablemetadata to
see this). The size of all live sstables are about 920 GiB.

Then there are tmp-la-*Data.db and tmplink-la-*Data.db files (the tmplink
files are hardlinks to the tmp file due to CASSANDRA-6916). I guess that
these come from the single active compaction. The total size of these files
are around ~65 GiB.

On the compaction side compactionstats shows that there's just one
compaction running, which is heavily CPU bound: (I've reformatted the
output here)

pending tasks: 5390
bytes done: 673623792733 (673 GiB)
bytes left: 3325656896682 (3325 GiB)
Active compaction remaining time :   2h44m39s

Why is the bytes done and especially bytes left such big? I don't have that
much data in my node.

Also how does Cassandra calculate the pending tasks with LCS?

Why are there a few such big SSTables in the active sstable list? Is it
because LCS falls back to STCS if L0 is too full? Should I use the
stcs_im_l0:false option? What will happen to these big sstables in the

I'm currently just waiting for the compactions to eventually finish, but
I'm hoping to learn in more detail what the system does and possibly to
help similar migration in the future.


 - Garo

View raw message