tephra-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andreas Neumann (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (TEPHRA-241) Introduce a way to limit the size of a transaction
Date Wed, 30 Aug 2017 23:24:02 GMT

    [ https://issues.apache.org/jira/browse/TEPHRA-241?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16148207#comment-16148207
] 

Andreas Neumann commented on TEPHRA-241:
----------------------------------------

Also, the client can check the size of the change set and avoid sending it over the wire if
it exceeds the max allowed size.

> Introduce a way to limit the size of a transaction
> --------------------------------------------------
>
>                 Key: TEPHRA-241
>                 URL: https://issues.apache.org/jira/browse/TEPHRA-241
>             Project: Tephra
>          Issue Type: Improvement
>          Components: api, manager
>    Affects Versions: 0.12.0-incubating
>            Reporter: Andreas Neumann
>            Assignee: Andreas Neumann
>             Fix For: 0.13.0-incubating
>
>
> When clients perform a huge number of writes in a short transaction, that can result
in huge change sets. For example, if a client performs 10M writes and sends that change set
over, that can easily be 1GB large. The transaction manager will keep this in memory. It will
also write this as an edit to the transaction log.
> Assume it runs out of memory because the change set is too large. It crashes and when
it restarts, it will replay the log, load that huge change set again, and crash again. 
> To prevent this kind of systemic failure, and to encourage developers to use long transactions
when performing many writes, we can introduce two new properties in the configuration:
> - change set warn threshold: if a change set exceeds this size, a warning is logged.

> - change set reject threshold: if a change set exceeds this size, it is rejected (canCommit
will throw an exception) and that will fail the transaction.
> Both thresholds should be Long.MAX_VALUE by default, to preserve existing behavior after
upgrade. 



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message