[
https://issues.apache.org/jira/browse/KUDU-2317?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16370646#comment-16370646
]
Todd Lipcon commented on KUDU-2317:
-----------------------------------
[~wdberkeley] do we have a JIRA for the root cause of why the metadata got so
large in the first place? a 50MB tablet meta means it's probably using tons of
memory for the metadata, plus any flush is rewriting the whole thing
> Tablet can fail to copy if it has metadata that is too large
> ------------------------------------------------------------
>
> Key: KUDU-2317
> URL: https://issues.apache.org/jira/browse/KUDU-2317
> Project: Kudu
> Issue Type: Bug
> Components: tablet copy
> Affects Versions: 1.6.0
> Reporter: Will Berkeley
> Priority: Major
>
> The BeginTabletCopySessionResponsePB contains a TabletSuperBlockPB, which
> contains a listing of all of a tablet's rowsets. If the number of rowsets is
> so large that the superblock exceeds the rpc max message size (default 50MB),
> it's not possible to copy the tablet.
> The copy client will log messages like
> {noformat}
> W0214 14:14:51.899601 13430 consensus_peers.cc:396] T
> 4e3b863d211545aea24a553d48ec3866 P 43f60e1982444bb58405ca3bdfbadcb4 -> Peer
> 95b8321262394992a03c334039b1fe5e (tserver-00:7050): Unable to begin Tablet
> Copy on peer: Network error: RPC frame had a length of 52536004, but we only
> support messages up to 52428800 bytes long.
> {noformat}
> The copy source will log messages like
> {noformat}
> W0214 14:14:51.965180 15179 serialization.cc:64] Serialized
> kudu.tserver.BeginTabletCopySessionResponsePB (58609169 bytes) is larger than
> the maximum configured RPC message size (52428800 bytes). Sending anyway, but
> peer may reject the data.
> {noformat}
> An additional problem this causes is that the connection between the source
> and client is torn down, which can cause other tablet copies to fail.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
