accumulo-notifications mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Josh Elser (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (ACCUMULO-3232) Improve consumption of WAL header in partial replication case
Date Tue, 14 Oct 2014 00:04:34 GMT

    [ https://issues.apache.org/jira/browse/ACCUMULO-3232?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14170271#comment-14170271
] 

Josh Elser commented on ACCUMULO-3232:
--------------------------------------

This is a little awkward because the API doesn't want to provide an {{FSDataInputStream}}
(because of how encryption on disk was implemented), but instead provides a {{DataInputStream}}.
For the case when the {{decryptingInputStream}} is the same as the {{originalInput}} in {{DFSLoggerInputStreams}},
we can up-cast the {{decryptingInputStream}} back to a {{FSDataInputStream}} and efficiently
seek to the offset in the WAL we want. If we only have the {{DfsInputStream}} (when encryption
is enabled), we have no option but to still read all of the records, although we can avoid
the overhead of constructing the {{LogFileKey}} and {{LogFileValue}} pairs and just read the
bytes.

> Improve consumption of WAL header in partial replication case
> -------------------------------------------------------------
>
>                 Key: ACCUMULO-3232
>                 URL: https://issues.apache.org/jira/browse/ACCUMULO-3232
>             Project: Accumulo
>          Issue Type: Improvement
>          Components: replication
>            Reporter: Josh Elser
>            Assignee: Josh Elser
>             Fix For: 1.7.0
>
>
> Consider a system that is actively replicating from one instance to another. Specifically,
assume there is one WAL that is currently being replicated to the destination and the source
instance is shutdown.
> When the source instance is restarted, it will notice that the WAL has read through N
{{LogFileKey}}/{{LogFileValue}} pairs (from before it was shutdown) and while proceed past
these records to get to the data in the file which it needs to read.
> We have to re-read each of these pairs from the file because the WAL is an append-only
structure, and we can't efficiently seek to some point in the file, as we wouldn't know how
to correlate the byte offset to entries.
> As we read the WAL, in addition (or perhaps instead of) tracking the offset in the WAL,
it would be good to track the correlation of N bytes read to M records consumed which would
help us better resume replication.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message