2

I recently ran YCSB benchmarks on CouchDB with 2 different workloads. Both workloads were using a database containing 500.000 documents and both of them executed 100.000 operations. The distribution of operations for each workload was:

Workload OLTP
READS=80%
INSERTS=10%
UPDATES=10%

Workload Cloud
READS=80%
INSERTS=20%
UPDATES=0%

Running both benchmarks i gathered the following results.

ycsb_results Now i don't understand why INSERT operations are being stable, while READs grow in latency the more READs are being executed. Shouldn't it be the other way around? At one point OLTP-READs even surpass the INSERTs. Why so?

Also, why are the READs in the second workload (Read Cloud) faster than in the first (OLTP Read)? The Cloud workload doesn't use updates, but how is this influencing the speed of READs?

Anton Horst
  • 263
  • 1
  • 4
  • 11
  • actually, since there is no client for couchDB in the original YCSB, you would have to show how you implemented update to get a definitive answer. – Asya Kamsky Mar 18 '15 at 00:43

1 Answers1

2

Check if CouchDB setting ‘delayed_commits’ is ‘true’. If it is – it may be an answer for your first question.

ermouth
  • 835
  • 7
  • 12
  • delayed_commits is set to 'true' in my current couchdb settings. So thats why inserts are being stable throughout the benchmark? – Anton Horst Jan 28 '15 at 21:26
  • It was just a clue – you better test it. For second question – I think you have ‘faster’ reads in OLTP scenarion, cause you, in fact, doing less read requests. Each UPDATE needs READ first – and INSERT does not. – ermouth Jan 31 '15 at 14:19