[
https://issues.apache.org/jira/browse/PHOENIX-8007?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tanuj Khurana updated PHOENIX-8007:
-----------------------------------
Description: PHOENIX-7727 added _phoenix.index.useServerMetadata_ (default
true) to eliminate the client _addServerCache_ RPC by having the region server
rebuild the I{_}ndexMaintainer{_} from its PTable cache per batch
{_}PhoenixIndexMetaDataBuilder.getIndexMetaDataCacheFromPTable{_}. It builds a
fresh PhoenixConnection per batch which contributes to wall latency and under
concurrent load has lock contention. I have attached both wall and lock
flamegraphs which show the contribution of new connection. I think we can
create the connection lazily per region but we have to be careful if the
connection is reaped and closed for example when cqsi closes. (was:
PHOENIX-7727 added `{_}phoenix.index.useServerMetadata{_}` (default
*{*}true{*}*) to eliminate the client `addServerCache` RPC by having the region
server rebuild the `IndexMaintainer` from its PTable cache per batch
(`PhoenixIndexMetaDataBuilder.getIndexMetaDataCacheFromPTable`). It builds a
fresh
`PhoenixConnection` per batch which contributes to wall latency and under
concurrent load has lock contention. I have attached both wall and lock
flamegraphs which show the contribution of new connection. I think we can
create the connection lazily per region but we have to be careful if the
connection is reaped and closed for example when cqsi closes.)
> Index metadata lookup on the server creates a new Server connection on every
> commit
> -----------------------------------------------------------------------------------
>
> Key: PHOENIX-8007
> URL: https://issues.apache.org/jira/browse/PHOENIX-8007
> Project: Phoenix
> Issue Type: Improvement
> Affects Versions: 5.4.0, 5.3.1
> Reporter: Tanuj Khurana
> Priority: Major
> Attachments: 7727-server-lock-flamegraph.html,
> 7727-server-wall-flamegraph.html
>
>
> PHOENIX-7727 added _phoenix.index.useServerMetadata_ (default true) to
> eliminate the client _addServerCache_ RPC by having the region server rebuild
> the I{_}ndexMaintainer{_} from its PTable cache per batch
> {_}PhoenixIndexMetaDataBuilder.getIndexMetaDataCacheFromPTable{_}. It builds
> a fresh PhoenixConnection per batch which contributes to wall latency and
> under concurrent load has lock contention. I have attached both wall and lock
> flamegraphs which show the contribution of new connection. I think we can
> create the connection lazily per region but we have to be careful if the
> connection is reaped and closed for example when cqsi closes.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
