In our cluster we have a columnfamiliy where its key is static and its columns have many inserts and deletes, like an inverse index. Over time, the read latency of keys with manys columns increase, now, its broke by timeouts.
We use ColumnSliceIterator with default count of 100 itens, slicing from Min to Max Characters. It's only work when pagination is less than 5.
We thought that tumbstones columns could be causing the problem, so, we reduced the gc_grace_seconds and we executed many repairs, but nothing changes.
Any idea?
Yes, tombstones will absolutely cause this.
You're on the right track with reducing gcgs, but the key to throwing away tombstones is compaction. You should consider switching to Leveled Compaction, which is much more aggressive about merging rows together: http://www.datastax.com/dev/blog/when-to-use-leveled-compaction
Incidentally, we've added request tracing to Cassandra 1.2 to make this kind of thing easier to diagnose: http://www.datastax.com/dev/blog/tracing-in-cassandra-1-2