Redis (Cluster) notes

Benchmark tool

Comes with Redis. Very good with variety of options. 

Server on AWS  (client on my local machine)

PS C:\Users\user> redis-benchmark -t set -n 100000 -h <MY_AWS_FREE_TIER_IP> -p 81

====== SET ======

  100000 requests completed in 40.93 seconds

  50 parallel clients

  3 bytes payload

  keep alive: 1

0.00% <= 14 milliseconds

0.22% <= 15 milliseconds

10.06% <= 16 milliseconds

24.73% <= 17 milliseconds

30.37% <= 18 milliseconds

32.67% <= 19 milliseconds

43.62% <= 20 milliseconds

63.83% <= 21 milliseconds

71.76% <= 22 milliseconds

74.68% <= 23 milliseconds

79.34% <= 24 milliseconds

89.36% <= 25 milliseconds

95.18% <= 26 milliseconds

97.42% <= 27 milliseconds

98.39% <= 28 milliseconds

98.87% <= 29 milliseconds

99.11% <= 30 milliseconds

99.36% <= 31 milliseconds

99.54% <= 32 milliseconds

99.65% <= 33 milliseconds

99.72% <= 34 milliseconds

99.80% <= 35 milliseconds

99.85% <= 36 milliseconds

99.89% <= 37 milliseconds

99.91% <= 38 milliseconds

99.93% <= 39 milliseconds

99.95% <= 40 milliseconds

99.96% <= 41 milliseconds

99.97% <= 42 milliseconds

99.98% <= 44 milliseconds

99.98% <= 45 milliseconds

100.00% <= 46 milliseconds

100.00% <= 50 milliseconds

100.00% <= 50 milliseconds

2443.38 requests per second

Localhost: (both client and server)

PS C:\Users\user> redis-benchmark -t set -n 100000

====== SET ======

  100000 requests completed in 1.42 seconds

  50 parallel clients

  3 bytes payload

  keep alive: 1

89.97% <= 1 milliseconds

99.88% <= 2 milliseconds

99.91% <= 3 milliseconds

99.93% <= 4 milliseconds

99.95% <= 8 milliseconds

99.96% <= 9 milliseconds

99.98% <= 10 milliseconds

99.99% <= 11 milliseconds

99.99% <= 12 milliseconds

99.99% <= 13 milliseconds

99.99% <= 14 milliseconds

99.99% <= 15 milliseconds

100.00% <= 16 milliseconds

100.00% <= 17 milliseconds

100.00% <= 18 milliseconds

100.00% <= 19 milliseconds

70521.86 requests per second

With pipelining (QPS is much higher but latency for 90 percentile requests is much higher)

PS C:\Users\user> redis-benchmark -t set -n 100000 -P 100

====== SET ======

  100000 requests completed in 0.21 seconds

  50 parallel clients

  3 bytes payload

  keep alive: 1

0.00% <= 1 milliseconds

0.60% <= 2 milliseconds

1.00% <= 3 milliseconds

1.10% <= 4 milliseconds

1.30% <= 5 milliseconds

1.40% <= 6 milliseconds

10.70% <= 7 milliseconds

35.80% <= 8 milliseconds

46.70% <= 9 milliseconds

57.00% <= 10 milliseconds

69.70% <= 11 milliseconds

79.10% <= 12 milliseconds

86.60% <= 13 milliseconds

91.70% <= 14 milliseconds

94.00% <= 15 milliseconds

97.00% <= 16 milliseconds

99.30% <= 17 milliseconds

100.00% <= 17 milliseconds

467289.72 requests per second

1. Redis is single threaded, can fork another thread(process?) for persistence.
2. If running in cluster mode, one node should have n/2 instances (master + slave) where n = NUM_CORE since one process for serving commands, another for persistence
3. RDB vs AOF persistence
4. Slave can be configured to become master if master hasn't been contacted in a while.
5. Total ~16K Hashslots.
6. Resharding (redistribution of keys) is always manual, whether adding a node or deleting one. Failover is automatic since slave already has the the same keys as master.
7. Pipelining will increase throughput but 90-95 percentile latency will be very high. Essentially 100 percentile latency will be lower than non-pipeline version but for other percentiles it will be very high.
8. Others : Recently Geo commands were added. Though tiles38 is also there for the similar stuff.