Hi Jeff,
Thanks for the information! That helps clear up a lot of things for me. We've had issues with SSTables not being dropped despite being past their TTLs (no read repairs or deletes) and I think issue 9572 explains why (definitely going to look into updating soon). Just to clarify, in the current version of Cassandra when do fully expired SSTables get dropped? Is it when a minor compaction runs or is it separate from minor compactions? Also thanks to the link to the slides, great stuff! Jerome ________________________________ From: Jeff Jirsa <jeff.ji...@crowdstrike.com> Sent: Friday, June 17, 2016 3:25:07 PM To: user@cassandra.apache.org Subject: Re: Understanding when Cassandra drops expired time series data First, DTCS in 2.0.15 has some weird behaviors - https://issues.apache.org/jira/browse/CASSANDRA-9572 . That said, some other general notes: Data deleted by TTL isn’t the same as issuing a delete – each expiring cell internally has a ttl/timestamp at which it will be converted into a tombstone. There is no tombstone added to the memtable, or flushed to disk – it just treats the expired cells as tombstones once they’re past that timestamp. Cassandra’s getFullyExpiredSSTables() will consider a table fully expired if (and only if) all cells within that table are expired (current time > max timestamp ) AND the sstable timestamps don’t overlap with others that aren’t fully expired. Björn talks about this in https://issues.apache.org/jira/browse/CASSANDRA-8243 - the intent here is so that explicit deletes (which do create tombstones) won’t be GC’d from an otherwise fully expired sstable if they’re covering data in a more recent sstable – without this check, we could accidentally bring dead data back to life. In an append only time series workload this would be unusual, but not impossible. Unfortunately, read repairs (foreground/blocking, if you write with CL < ALL and read with CL > ONE) will cause cells written with old timestamps to be written into the newly flushed sstables, which creates sstables with wide gaps between minTimestamp and maxTimestamp (you could have a read repair pull data that is 23 hours old into a new sstable, and now that one sstable spans 23 hours, and isn’t fully expired until the oldest data is 47 hours old). There’s an open ticket (https://issues.apache.org/jira/browse/CASSANDRA-10496 ) meant to make this behavior ‘better’ in the future by splitting those old read-repaired cells from the newly flushed sstables. I gave a talk on a lot of this behavior last year at Summit (http://www.slideshare.net/JeffJirsa1/cassandra-summit-2015-real-world-dtcs-for-operators ) - if you’re running time series in production on DTCS, it’s worth a glance. From: jerome <jeromefroel...@hotmail.com> Reply-To: "user@cassandra.apache.org" <user@cassandra.apache.org> Date: Friday, June 17, 2016 at 11:52 AM To: "user@cassandra.apache.org" <user@cassandra.apache.org> Subject: Understanding when Cassandra drops expired time series data Hello! Recently I have been trying to familiarize myself with Cassandra but don't quite understand when data is removed from disk after it has been deleted. The use case I'm particularly interested is expiring time series data with DTCS. As an example, I created the following table: CREATE TABLE metrics ( metric_id text, time timestamp, value double, PRIMARY KEY (metric_id, time), ) WITH CLUSTERING ORDER BY (time DESC) AND default_time_to_live = 86400 AND gc_grace_seconds = 3600 AND compaction = { 'class': 'DateTieredCompactionStrategy', 'timestamp_resolution':'MICROSECONDS', 'base_time_seconds':'3600', 'max_sstable_age_days':'365', 'min_threshold':'4' }; I understand that Cassandra will create a tombstone for all rows inserted into this table 24 hours after they are inserted (86400 seconds). These tombstones will first be written to an in-memory Memtable and then flushed to disk as an SSTable when the Memtable reaches a certain size. My question is when will the data that is now expired be removed from disk? Is it the next time the SSTable which contains the data gets compacted? So, with DTCS and min_threshold set to four, we would wait until at least three other SSTables are in the same time window as the expired data, and then those SSTables will be compacted into a SSTable without the expired data. Is it only during this compaction that the data will be removed? It seems to me that this would require Cassandra to maintain some metadata on which rows have been deleted since the newer tombstones would likely not be in the older SSTables that are being compacted. Also, I'm aware that Cassandra can drop entire SSTables if they contain only expired data but I'm unsure of what qualifies as expired data (is it just SSTables whose maximum timestamp is past the default TTL for the table?) and when such SSTables are dropped. Alternatively, do the SSTables which contain the tombstones have to be compacted with the SSTables which contain the expired data for the data to be removed? It seems to me that this could result in Cassandra holding the expired data long after it has expired since it's waiting for the new tombstones to be compacted with the older expired data. Finally, I was also unsure when the tombstones themselves are removed. I know Cassandra does not delete them until after gc_grace_seconds but it can't delete the tombstones until it's sure the expired data has been deleted right? Otherwise it would see the expired data as being valid. Consequently, it seems to me that the question of when tombstones are deleted is intimately tied to the questions above. Thanks in advance! If it helps I've been experimenting with version 2.0.15 myself.
