[
https://issues.apache.org/jira/browse/CASSANDRA-8243?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14210057#comment-14210057
]
Björn Hegerfors commented on CASSANDRA-8243:
--------------------------------------------
An expired column is equivalent to a tombstone with the same timestamp in
Cassandra's eyes, right? Compactions even turn them into tombstones, if they
can't be immediately purged. So to simplify, we're dealing with all-tombstone
SSTables. Both the old and new implementation agree that removing an SSTable
can only happen if the oldest SSTable (the one with lowest minTimestamp) is
all-tombstones (= has fully expired). Both implementations also agree that this
oldest SSTable may not overlap (in time span) with an SSTable containing any
non-tombtone data. If there is no such overlap, everything in any SSTable (with
an overlapping row range, anyway) written with a timestamp less than or equal
to this oldest table's maxTimestamp is guaranteed to be a tombstone.
Also, since any SSTable that either of the implementations remove is an
all-tombstone SSTable, the only thing that can happen is that something is
resurrected. Combined with the reasoning in my previous paragraph, the only
thing that could be resurrected when a tombstone for column x with timestamp t
is removed is another tombstone for column x, with a lower timestamp t'! When
could that matter? Only if some other SSTable makes a constructive write to
column x in the interval (t', t]. But that's impossible, because that would
then be an SSTable containing some non-tombstone data with a minTimestamp less
than or equal to the oldest SSTable's maxTimestamp, which goes against the
assumption that no such SSTable exists!
There you have a proof by contradiction that the oldest SSTable can be safely
removed if it is all-tombstones and doesn't overlap with any SSTable containing
any non-tombstone data. If we then consider the oldest SSTable free to remove,
the same rules apply to the oldest remaining SSTable and so on. This is the
rule that my implementation uses. From the comments it looks like we already
agree intuitively on this, but I though a more formal proof like this might
help this get committed. [~slebresne] any reason to still not submit this patch
to 2.0?
Oh, and I noticed that I didn't update the Javadoc, so here comes a new patch.
> DTCS can leave time-overlaps, limiting ability to expire entire SSTables
> ------------------------------------------------------------------------
>
> Key: CASSANDRA-8243
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8243
> Project: Cassandra
> Issue Type: Bug
> Reporter: Björn Hegerfors
> Assignee: Björn Hegerfors
> Priority: Minor
> Labels: compaction, performance
> Fix For: 2.0.12, 2.1.3
>
> Attachments: cassandra-trunk-CASSANDRA-8243-aggressiveTTLExpiry.txt,
> cassandra-trunk-CASSANDRA-8243-aggressiveTTLExpiry.txt
>
>
> CASSANDRA-6602 (DTCS) and CASSANDRA-5228 are supposed to be a perfect match
> for tables where every value is written with a TTL. DTCS makes sure to keep
> old data separate from new data. So shortly after the TTL has passed,
> Cassandra should be able to throw away the whole SSTable containing a given
> data point.
> CASSANDRA-5228 deletes the very oldest SSTables, and only if they don't
> overlap (in terms of timestamps) with another SSTable which cannot be deleted.
> DTCS however, can't guarantee that SSTables won't overlap (again, in terms of
> timestamps). In a test that I ran, every single SSTable overlapped with its
> nearest neighbors by a very tiny amount. My reasoning for why this could
> happen is that the dumped memtables were already overlapping from the start.
> DTCS will never create an overlap where there is none. I surmised that this
> happened in my case because I sent parallel writes which must have come out
> of order. This was just locally, and out of order writes should be much more
> common non-locally.
> That means that the SSTable removal optimization may never get a chance to
> kick in!
> I can see two solutions:
> 1. Make DTCS split SSTables on time window borders. This will essentially
> only be done on a newly dumped memtable once every base_time_seconds.
> 2. Make TTL SSTable expiry more aggressive. Relax the conditions on which an
> SSTable can be dropped completely, of course without affecting any semantics.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
