[
https://issues.apache.org/jira/browse/IGNITE-6409?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vladimir Ozerov updated IGNITE-6409:
------------------------------------
Description:
Currently the only way to update standard user index is through H2
infrastructure. See {{GridH2IndexBase#put}} and it's usages. For every update
we have to construct a kind of H2 row which will be passed to index. This
operation might require unnecessary memory copying and deserializations. We
should try to find a way to bypass H2 stuff altogether and add data to
underlying {{BPlusTree}} directly.
Here is how this could be achieved:
1) Actual tree is {{H2Tree}}. It is accompanied by several helper classes
({{InlineIndexHelper}}, {{H2*IO}}).
2) We should remove all H2 stuff from there, and start using {{CacheDataRow}}
instead of {{SearchRow}}
3) Refactored classes should be moved to core module.
4) Conversion to {{SearchRow}} should happen inside {{H2TreeIndex}} only when
it is really needed.
5) Two operations {{putNative}} and {{getNative}} should be added to
{{GridH2IndexBase}}. They should accept normal key-value objects instead of H2
rows.
6) Finally, all index index update routines (cache put/get, {{CREATE INDEX}})
should be moved to {{putNative}}/{{getNative}} methods.
Note that IO classes should be reworked carefully to maintain backward
compatibility.
was:
Currently the only way to update standard user index is through H2
infrastructure. See {{GridH2IndexBase#put}} and it's usages. For every update
we have to construct a kind of H2 row which will be passed to index. This
operation might require unnecessary memory copying and deserializations. We
should try to find a way to bypass H2 stuff altogether and add data to
underlying {{BPlusTree}} directly.
Here is how this could be achieved:
1) Actual tree is {{H2Tree}}. It is accompanied by several helper classes
({{InlineIndexHelper}}, {{H2*IO}}).
2) We should remove all H2 stuff from there, and start using {{CacheDataRow}}
instead of {{SearchRow}}
3) Conversion to {{SearchRow}} should happen inside {{H2TreeIndex}} only when
it is really needed.
4) Two operations {{putNative}} and {{getNative}} should be added to
{{GridH2IndexBase}}. They should accept normal key-value objects instead of H2
rows.
5) Finally, index creation routine should be moved to
{{putNative}}/{{getNative}} methods.
Note that IO classes should be reworked carefully to maintain backward
compatibility.
> SQL: bypass H2 during index update
> ----------------------------------
>
> Key: IGNITE-6409
> URL: https://issues.apache.org/jira/browse/IGNITE-6409
> Project: Ignite
> Issue Type: Task
> Components: persistence, sql
> Affects Versions: 2.1
> Reporter: Vladimir Ozerov
> Labels: iep-1, performance
> Fix For: 2.4
>
>
> Currently the only way to update standard user index is through H2
> infrastructure. See {{GridH2IndexBase#put}} and it's usages. For every update
> we have to construct a kind of H2 row which will be passed to index. This
> operation might require unnecessary memory copying and deserializations. We
> should try to find a way to bypass H2 stuff altogether and add data to
> underlying {{BPlusTree}} directly.
> Here is how this could be achieved:
> 1) Actual tree is {{H2Tree}}. It is accompanied by several helper classes
> ({{InlineIndexHelper}}, {{H2*IO}}).
> 2) We should remove all H2 stuff from there, and start using {{CacheDataRow}}
> instead of {{SearchRow}}
> 3) Refactored classes should be moved to core module.
> 4) Conversion to {{SearchRow}} should happen inside {{H2TreeIndex}} only when
> it is really needed.
> 5) Two operations {{putNative}} and {{getNative}} should be added to
> {{GridH2IndexBase}}. They should accept normal key-value objects instead of
> H2 rows.
> 6) Finally, all index index update routines (cache put/get, {{CREATE INDEX}})
> should be moved to {{putNative}}/{{getNative}} methods.
> Note that IO classes should be reworked carefully to maintain backward
> compatibility.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
