[
https://issues.apache.org/jira/browse/CASSANDRA-9420?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14552266#comment-14552266
]
Sylvain Lebresne commented on CASSANDRA-9420:
---------------------------------------------
bq. What we should have done back then was to make that option not a default,
but an enforced constrain. So that every update/insert always comes with the
same TTL
The problem Björn wants to improve is explicitly the case where there is "lots
of variations in TTLs", so such option won't really help. And as he said, when
all TTLs are the same, DTCS is able to drop expired sstable very efficiently so
that this kind of option is not really needed for that case either.
bq. Even if gc gs is 0, we must ensure that there is no intersection with other
sstables, before purging it.
Yes, but as the description mentions time series and DTCS, so I assumed the
"min timestamp" checks would solve this. But I was wrong. As Björn says, with
variable TTLs, the "min timestamp" check will rarely help.
That said, there is probably better heuristics to improve this case. For
instance, in theory, we could use the the sstable max/min clustering values
which, for time series with DTCS, would allow to decide that most cells can't
intersect any non-compacted sstable (one question being how efficient we can
make that check be).
Because Cassandra cannot guarantee you won't override a column, so adding an
option that tells it to assume it blindly feels pretty dangerous and I'm
personally really not a fan.
> Table option for promising that you will never touch a column twice
> -------------------------------------------------------------------
>
> Key: CASSANDRA-9420
> URL: https://issues.apache.org/jira/browse/CASSANDRA-9420
> Project: Cassandra
> Issue Type: New Feature
> Reporter: Björn Hegerfors
>
> There are time series use cases where you write all values with various TTLs,
> have GC grace = 0 and never ever update or delete a column after insertion.
> In the case where all TTLs are the same, DTCS with recent patches works
> great. But when there is lots of variations in TTLs, you are forced to choose
> between splitting your table into multiple TTL tiers or having your SSTables
> filled to the majority with tombstones. Or running frequent major compactions.
> The problem stems from the fact that Cassandra plays safe when a TTL has
> expired, and turns it into a tombstone, rather than getting rid of it on the
> spot. The reason is that this TTL _may_ have been in a column which has had
> an earlier write without (or with a higher) TTL. And then that one should now
> be deleted too.
> I propose that there should be table level setting to say "I guarantee that
> there will never be any updates to any columns". The effect of enabling that
> option is that all tombstones and expired TTLs should always be immediately
> removed during compaction. And the check for dropping entirely expired
> SSTables can be very loosened for these tables.
> This option should probably require gc_grace_seconds to be set to zero. It's
> also questionable if writes without TTL should be allowed to such a table,
> since those would become constants.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
