Change data capture (CDC) provides a mechanism to flag specific tables for archival as well as rejecting writes to those
tables once a configurable size-on-disk for the CDC log is reached. An operator can enable CDC on a table by setting the
cdc=true (either when creating the table or
altering it). Upon CommitLogSegment creation, a hard-link to the segment is created in the
directory specified in
cassandra.yaml. On segment fsync to disk, if CDC data is present anywhere in the segment a
<segment_name>_cdc.idx file is also created with the integer offset of how much data in the original segment is persisted
to disk. Upon final segment flush, a second line with the human-readable word “COMPLETED” will be added to the _cdc.idx
file indicating that Cassandra has completed all processing on the file.
We we use an index file rather than just encouraging clients to parse the log realtime off a memory mapped handle as data can be reflected in a kernel buffer that is not yet persisted to disk. Parsing only up to the listed offset in the _cdc.idx file will ensure that you only parse CDC data for data that is durable.
A threshold of total disk space allowed is specified in the yaml at which time newly allocated CommitLogSegments will not allow CDC data until a consumer parses and removes files from the specified cdc_raw directory.
CDC is enable or disable through the cdc table property, for instance:
CREATE TABLE foo (a int, b text, PRIMARY KEY(a)) WITH cdc=true; ALTER TABLE foo WITH cdc=true; ALTER TABLE foo WITH cdc=false;
The following cassandra.yaml are available for CDC:
cdc_free_space_in_mb: (default: min of 4096 and 1/8th volume space)
cdc_raw_directoryto prevent burning CPU cycles unnecessarily. Default is to check 4 times per second.
Do not enable CDC without some kind of consumption process in-place.
If CDC is enabled on a node and then on a table, the
cdc_free_space_in_mb will fill up and then writes to
CDC-enabled tables will be rejected unless some consumption process is in place.