[
https://issues.apache.org/jira/browse/CASSANDRA-6916?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13963040#comment-13963040
]
Marcus Eriksson commented on CASSANDRA-6916:
--------------------------------------------
In general the patch looks good, a few comments;
* Can we get some benchmarks done if this is worth it? I know intuitively it
feels like it would be a big win, not only for the page cache, but also since
we get to only read one sstable for the compacted keys, but still would be nice
to have some evidence. I guess we could also do this to find a good default
value for sstable_preemptive_open_interval_in_mb
* Problem with fully expired sstables in CompactionTask, we need to replace
"toCompact" instead of "actuallyCompact".
* KeyCacheTest fails (not had time to look deeper into it)
* A few comments around the replaceReader/replaceCompacted... and related in
CompactionTask would be helpful, a bit harder to grasp now that we start using
the compacted sstables earlier and then call DataTracker.replace(...) with an
empty list.
* In SSTR.cloneWithNewStart(..), CLibrary.trySkipCache(dfile.path, 0, 0); is
called when newStart > last, is that on purpose? I guess we could drop the page
cache for the whole file in this case?
* In MMappedSegmentedFile, line 199, it looks like that ternary expression is
not really needed since we add raf.length() in boundaries above?
* Preheat comment left in CompactionTask
We could probably do the same thing for the rest of the compactions
(upgradesstables, scrub, etc)? Lets do that in a followup ticket (if at all)
> Preemptive opening of compaction result
> ---------------------------------------
>
> Key: CASSANDRA-6916
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6916
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Benedict
> Assignee: Benedict
> Priority: Minor
> Fix For: 2.1
>
>
> Related to CASSANDRA-6812, but a little simpler: when compacting, we mess
> quite badly with the page cache. One thing we can do to mitigate this problem
> is to use the sstable we're writing before we've finished writing it, and to
> drop the regions from the old sstables from the page cache as soon as the new
> sstables have them (even if they're only written to the page cache). This
> should minimise any page cache churn, as the old sstables must be larger than
> the new sstable, and since both will be in memory, dropping the old sstables
> is at least as good as dropping the new.
> The approach is quite straight-forward. Every X MB written:
> # grab flushed length of index file;
> # grab second to last index summary record, after excluding those that point
> to positions after the flushed length;
> # open index file, and check that our last record doesn't occur outside of
> the flushed length of the data file (pretty unlikely)
> # Open the sstable with the calculated upper bound
> Some complications:
> # must keep running copy of compression metadata for reopening with
> # we need to be able to replace an sstable with itself but a different lower
> bound
> # we need to drop the old page cache only when readers have finished
--
This message was sent by Atlassian JIRA
(v6.2#6252)
