Shubham Roy created PHOENIX-8032:
------------------------------------
Summary: Support SPLIT ON in CREATE CDC to pre-split the CDC index
Key: PHOENIX-8032
URL: https://issues.apache.org/jira/browse/PHOENIX-8032
Project: Phoenix
Issue Type: Improvement
Reporter: Shubham Roy
Assignee: Shubham Roy
Fix For: 5.4.0, 5.3.3
h3. Problem
{{CREATE TABLE}} and {{CREATE INDEX}} accept {{SPLIT ON}}, but {{CREATE CDC}}
does not. The uncovered index behind a CDC is always created with a single
region, so every CDC index write lands on that one region until it splits by
size. On large, write-heavy tables, that region becomes a hotspot as soon as
CDC is enabled.
h3. Proposal
Accept an optional {{SPLIT ON}} clause after the CDC properties, in the same
position as in {{CREATE TABLE}} and {{CREATE INDEX}}, and use it to pre-split
the CDC index.
{code:sql}
CREATE CDC [IF NOT EXISTS] cdc_name ON table_ref
[INCLUDE (change_scope [, ...])]
[property [, ...]]
[SPLIT ON (split_point [, ...])]
-- example
CREATE CDC my_cdc ON my_table SPLIT ON ('4', '8', 'c')
{code}
* Split points apply to the CDC index row key: {{PARTITION_ID()}}—the data
table region's encoded name, consisting of 32 lowercase hexadecimal
characters—followed by {{PHOENIX_ROW_TIMESTAMP()}} and the data table primary
key. Multi-tenant tables put {{TENANT_ID}} first.
* Split points can be constants or bind parameters, as in {{CREATE TABLE}}.
Bind handling pre-splits the index using exactly those bytes.
* A CDC on a view keeps its index in the shared view index table, which does
not accept user-supplied split points. The clause is accepted there but has no
effect, matching the behavior of {{CREATE INDEX ... SPLIT ON}} on a view.
h3. Testing
* {{QueryParserTest#testCreateCDCWithSplitPoints}}: Verify parsing, clause
order, and bind count.
* {{CDCDefinitionIT#testCreateCDCWithSplitPoints}}: Verify that the CDC index
regions match a mix of literal and raw-byte bind split points.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
