And actually, full repair with 3.0/3.x would have the same effect (anticompation) unless you're using subrange repair.
On Mon, Nov 13, 2017 at 3:28 PM Jeff Jirsa <jji...@gmail.com> wrote: > Running incremental repair puts sstables into a “repaired” set (and an > unrepaired set), which results in something similar to what you’re > describing. > > Were you running / did you run incremental repair ? > > > -- > Jeff Jirsa > > > On Nov 13, 2017, at 5:04 AM, Nicolas Guyomar <nicolas.guyo...@gmail.com> > wrote: > > Hi everyone, > > I'm facing quite a strange behavior on STCS on 3.0.13, the strategy seems > to have "forgotten" about old sstables, and started a completely new cycle > from scratch, leaving the old sstables on disk untouched : > > Something happened on Nov 10 on every node, which resulted in all those > sstables left behind : > > -rw-r--r--. 8 cassandra cassandra 15G Nov 9 22:22 mc-4828-big-Data.db > -rw-r--r--. 8 cassandra cassandra 4.8G Nov 10 01:39 mc-4955-big-Data.db > -rw-r--r--. 8 cassandra cassandra 2.4G Nov 10 01:45 mc-4957-big-Data.db > -rw-r--r--. 8 cassandra cassandra 662M Nov 10 01:47 mc-4959-big-Data.db > -rw-r--r--. 8 cassandra cassandra 2.8G Nov 10 03:46 mc-5099-big-Data.db > -rw-r--r--. 8 cassandra cassandra 4.6G Nov 10 03:58 mc-5121-big-Data.db > -rw-r--r--. 7 cassandra cassandra 53M Nov 10 08:45 mc-5447-big-Data.db > -rw-r--r--. 7 cassandra cassandra 219M Nov 10 08:46 mc-5454-big-Data.db > -rw-r--r--. 7 cassandra cassandra 650M Nov 10 08:46 mc-5452-big-Data.db > -rw-r--r--. 7 cassandra cassandra 1.2G Nov 10 08:48 mc-5458-big-Data.db > -rw-r--r--. 7 cassandra cassandra 1.5G Nov 10 08:50 mc-5465-big-Data.db > -rw-r--r--. 7 cassandra cassandra 504M Nov 10 09:39 mc-5526-big-Data.db > -rw-r--r--. 7 cassandra cassandra 57M Nov 10 09:40 mc-5527-big-Data.db > -rw-r--r--. 7 cassandra cassandra 101M Nov 10 09:41 mc-5532-big-Data.db > -rw-r--r--. 7 cassandra cassandra 86M Nov 10 09:41 mc-5533-big-Data.db > -rw-r--r--. 7 cassandra cassandra 134M Nov 10 09:42 mc-5537-big-Data.db > -rw-r--r--. 7 cassandra cassandra 3.9G Nov 10 09:54 mc-5538-big-Data.db > *-rw-r--r--. 7 cassandra cassandra 1.3G Nov 10 09:57 mc-5548-big-Data.db* > -rw-r--r--. 6 cassandra cassandra 16G Nov 11 01:23 mc-6474-big-Data.db > -rw-r--r--. 4 cassandra cassandra 17G Nov 12 06:44 mc-7898-big-Data.db > -rw-r--r--. 3 cassandra cassandra 8.2G Nov 12 13:45 mc-8226-big-Data.db > -rw-r--r--. 2 cassandra cassandra 6.8G Nov 12 22:38 mc-8581-big-Data.db > -rw-r--r--. 2 cassandra cassandra 6.1G Nov 13 03:10 mc-8937-big-Data.db > -rw-r--r--. 2 cassandra cassandra 3.1G Nov 13 04:12 mc-9019-big-Data.db > -rw-r--r--. 2 cassandra cassandra 3.0G Nov 13 05:56 mc-9112-big-Data.db > -rw-r--r--. 2 cassandra cassandra 1.2G Nov 13 06:14 mc-9138-big-Data.db > -rw-r--r--. 2 cassandra cassandra 1.1G Nov 13 06:27 mc-9159-big-Data.db > -rw-r--r--. 2 cassandra cassandra 1.2G Nov 13 06:46 mc-9182-big-Data.db > -rw-r--r--. 1 cassandra cassandra 1.9G Nov 13 07:18 mc-9202-big-Data.db > -rw-r--r--. 1 cassandra cassandra 353M Nov 13 07:22 mc-9207-big-Data.db > -rw-r--r--. 1 cassandra cassandra 120M Nov 13 07:22 mc-9208-big-Data.db > -rw-r--r--. 1 cassandra cassandra 100M Nov 13 07:23 mc-9209-big-Data.db > -rw-r--r--. 1 cassandra cassandra 67M Nov 13 07:25 mc-9210-big-Data.db > -rw-r--r--. 1 cassandra cassandra 51M Nov 13 07:25 mc-9211-big-Data.db > -rw-r--r--. 1 cassandra cassandra 73M Nov 13 07:27 mc-9212-big-Data.db > > > TRACE logs for the Compaction Manager shows that sstables before Nov 10 > are grouped in different buckets than the one after Nov 10. > > At first I thought off some coldness behavior that would filter those > "old" sstables, but looking at the code > https://github.com/apache/cassandra/blob/cassandra-3.0/src/java/org/apache/cassandra/db/compaction/SizeTieredCompactionStrategy.java#L237 > I don't see any coldness or time pattern used to create bucket. > > I tried restarting the node but the buckets are still grouping in 2 > "groups" splitted around Nov 10 > > I may have missed sthg from the logs, but they are clear from error/warn > at that Nov 10 time > > For what it's worth, restarting the node fixed nodetool status from > reporting a wrong Load (nearly 2TB per node instead à 300Gb) => we are > loading some data for a week now, it seems that this can happen sometimes > > If anyone ever experienced that kind of behavior I'd be glad to know > whether it is OK or not, I'd like to avoid manually triggering JMX > UserDefinedCompaction ;) > > Thank you > > -- ----------------- Alexander Dejanovski France @alexanderdeja Consultant Apache Cassandra Consulting http://www.thelastpickle.com
