Cassandra Anti-Patterns
12 likes4,887 views
This document summarizes several Cassandra anti-patterns including: - Using a non-Oracle JVM which is not recommended. - Putting the commit log and data directories on the same disk which can impact performance. - Using EBS volumes on EC2 which can have unpredictable performance and throughput issues. - Configuring overly large JVM heaps over 16GB which can cause garbage collection issues. - Performing large batch mutations in a single operation which risks timeouts if not broken into smaller batches.
Ad
More Related Content
What's hot (19)
Viewers also liked (20)
Ad
Similar to Cassandra Anti-Patterns (20)
Ad
Recently uploaded (20)
Cassandra Anti-Patterns
- 1. Cassandra Anti-Patterns (in 5m) Matthew F. Dennis // @mdennis
- 2. Non-Sun (err, Non-Oracle) JVM ● No OpenJDK ● No Blackdown (anyone still use this?) ● Etc, etc, etc; just use the Sun (Oracle) JVM ● At least u22, but in general the latest release (unless you have specific reasons otherwise)
- 3. CommitLog+Data On The Same Disk ● Don't put the commit log and data directories on the same set of spindles – commit log gets a single spindle entirely to itself (standard consumer SATA disks easily sustain > 80 MB/s in sequential writes) ● DOES NOT APPLY TO SSDS or EC2 ● SSDs have no seek time ● EC2 ephemeral drives are still virtualized (but not the same as EBS) ● On EC2 or SSDs: use one RAID set for both the commit log and data directories
- 4. EBS volumes on EC2 ● Sounds great, nice feature set, but … ● Not predictable ● “freezes” are common ● Throughput limited in many cases ● Use ephemeral drives instead ● Stripe them ● Both commit log and data directory on the same raid set
- 5. Oversized JVM heaps ● 6 – 8 GB is good (assuming sufficient ram on your boxen) ● 10 – 12 GB is possible and in some circumstances “correct” ● 16GB == max JVM heap size ● > 16GB => badness ● JVM heap ~= boxen RAM => badness (always)
- 6. JVM heap size -v- GC suckage ~16GB GC Suckage ~10GB ~6GB JVM heap size
- 7. Large batch mutations (large in number of distinct rows) ● Timeout / failure => entire mutation must be retried => wasted work ● Larger mutations => higher likely hood of timehood ● 1000 mutations to perform? Do 100 batches of 10 in parallel instead of one batch of 1000 ● Exact number or rows/batch is variable depending on HW, network, load, etc; experiment! (10-100 is a good starting point)
- 8. OPP / BOP partitioner ● You probably shouldn't use it ● No really, you almost certainly shouldn't use it ● Creates hot spots ● Requires “baby sitting” from ops ● Not as well tested nor is it widely deployed
- 9. C* auto selection of tokens ● Always specify your initial token. ● Auto select doesn't do what you think it does nor does it do what you want – loadbalance is even worse, it doesn't currently do what you think, what you want or what it claims; “F#@* my cluster” would be a much more apt name than “loadbalance” – Future (next?) release of OPSC will remove your balancing woes
- 10. Super Columns ● 10 – 15 percent performance penalty on reads and writes ● Easier / better to use to composite columns – 0.8.x makes this a lot easier – Done manually in 0.7.x and is still better ● Devs working in C* code despise (loathe?) them ● API probably won't be deprecated, but implementation will be replaced behind the seen with composites (may be “ok” at that point to use them, but should probably just use composite API direclty) ● Cassandra and DataStax is committed to maintain the API going forward, even if the implementation changes
- 11. Read Before Write ● Race conditions ● Abuses/Thrashes cache (row, key and page) ● Increases latency ● Increases IO requirements (by a lot) ● Increases size in the client
- 12. Winblows ● Try to avoid it, you'll be happier – Not always possible? Then, “I'm sorry for your pain” ● Run 'nix (in particular, probably Linux) ● Easier to get help (IRC, email, meetups, etc) ● C* performs better ● Better tested ● Cheaper ● Wider deployed (by a lot)
- 13. Q? Cassandra Anti-Patterns Matthew F. Dennis // @mdennis