3 ms·
Understanding the Causes of Consistency Anomalies in Apache Cassandra [pdf]
- alexnewman 12y ago5ms is a really long time
- arielweisberg 12y agoI think that measuring propagation delay at CL.ONE isn't very interesting because if you care about propagation delay you also don't use CL.ONE. That said my passion is to get everything in C* off heap and get GC pauses down to single digit milliseconds. Old-gen pauses a handful of times a day with pause times 30 milliseconds or so and also nothing to write home about since it should be possible to a small old gen. All with good old CMS or G1. From what I know it's a pretty attainable goal. The main bad actor is memtables and fast efficient off-heap memory management for memtables is easy because you can throw away all associated memory when the memtable is flushed. That still wouldn't make using CL.ONE any better of an idea. GC pauses are one of several sources of propagation delay.
- _benedict 12y ago"For example, in Cassandra such internal activities include the flush, also referred as minor compaction, which flushes in-memory data from the memtable to an SSTable file on disk, and may be accompanied by a substantial GC since a large chunk of memory becomes available for recycling." So these researchers don't unfortunately understand the very basics of GC (or C* for that matter). The basic findings of their paper is to be expected given the question they asked, but the specifics of their findings should be considered suspect given the clear lack of understanding of the technologies they're interrogating.