Database Reference
In-Depth Information
slow Cassandra down if you make it read more than is necessary. This becomes especially
important if you frequently read partial rows.
in_memory_compaction_limit_in_mb
in_memory_compaction_limit_in_mb
This value represents the size limit for rows being compacted in memory. If a row is larger
than this, it will be staged onto disk and compacted in a slower, two-step process. The set-
ting's value is set to 64MB by default. Prior to Cassandra version 0.7, this was called the
row_warning_threshold_in_mb
.
gc_grace_seconds
gc_grace_seconds
This indicates the amount of time in seconds to wait before garbage-collecting tombstones.
Tombstones are markers indicating that a value was deleted, but they persist until this
threshold is met. By default it is set to 864,000, which is 10 days. This might seem like an in-
ordinately long time, but you have to allow enough time for tombstones to propagate through
the entire cluster to any replicas. That may still seem like a long time, but the idea is that you
could have hardware failures, and you'll need time to get those nodes back up and operation-
al so that tombstones can replicate to them too. If a downed node that did not receive tomb-
stone notifications comes back up after the tombstones have been truly deleted, read repair
will allow the untombstoned, undeleted data to slowly be introduced back into the cluster.
Of course, if you do a lot of deleting, you could tune this setting back for the sake of tidiness,
but it's probably not going to make a huge difference.
phi_convict_threshold
phi_convict_threshold
This marks the Phi value that must be reached in order for Cassandra to conclude that the
server is down. By default, this is set to 8, and in general you will not want to change this
setting. However, if your network is not performing optimally or you have other reasons to
suspect that Cassandra could get a false read on this, you can increase this setting to allow
more leniency in determining whether a node is actually down.
