check if datatype or sc, skip del operation in put path

[zeeshanlakhani]

Fixes https://bashoeng.atlassian.net/browse/RIAK-1504 (#452).

Initial comments/PR #452.

Improves performance by avoiding an extra delete operation to remove siblings through Solr on the put path when riak datatypes or strong consistency is used. Initial patch/issue was discovered by @drewkerrigan, https://github.com/basho-labs/yokozuna_perf_patch.

[zeeshanlakhani]

@seancribbs benchmarks:

local runs

The consolidated fix seems to work well.

Yokozuna Develop - Sets

Mean ops/s for put-check/before-load-fruit: 574.530657
Mean ops/s for put-check/current: 574.530657

The mean median latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 51.932547
The mean median latency for put-check/current/load-fruit-pb_latencies.csv: 51.932547

The mean 95th latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 103.327948
The mean 95th latency for put-check/current/load-fruit-pb_latencies.csv: 103.327948

Yokozuna Perf Put Path CRDTS w/ Fix - Sets

Mean ops/s for put-check/before-load-fruit: 682.978673
Mean ops/s for put-check/current: 682.978673

The mean median latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 44.457306
The mean median latency for put-check/current/load-fruit-pb_latencies.csv: 44.457306

The mean 95th latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 81.066028
The mean 95th latency for put-check/current/load-fruit-pb_latencies.csv: 81.066028

Yokozuna - Develop - Maps

Mean ops/s for put-check/before-load-fruit: 647.997282
Mean ops/s for put-check/current: 647.997282

The mean median latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 45.288336
The mean median latency for put-check/current/load-fruit-pb_latencies.csv: 45.288336

The mean 95th latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 93.365178
The mean 95th latency for put-check/current/load-fruit-pb_latencies.csv: 93.365178

Yokozuna Perf Put Path CRDTS w/ Fix - Maps

Mean ops/s for put-check/before-load-fruit: 696.033863
Mean ops/s for put-check/current: 696.033863

The mean median latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 43.415865
The mean median latency for put-check/current/load-fruit-pb_latencies.csv: 43.415865

The mean 95th latency for put-check/before-load-fruit/load-fruit-pb_latencies.csv: 80.376965
The mean 95th latency for put-check/current/load-fruit-pb_latencies.csv: 80.376965

Here's Drew's work that he did initially on some powerful machines and just removing the delete operation altogether:

Drew’s BB work - PrePatch

Mean ops/s for 20150121_214413: 2654.752636

The mean median latency for 20150121_214413/post-map-targets-json5-v-json-h_latencies.csv: 39.811105

The mean 95th latency for 20150121_214413/post-map-targets-json5-v-json-h_latencies.csv: 142.582000

Drew’s BB work - PostPatch (Full removal of Delete Op to Solr)

Mean ops/s for 20150121_221042: 6322.768601

The mean median latency for 20150121_221042/post-map-targets-json5-v-json-h_latencies.csv: 39.639982

The mean 95th latency for 20150121_221042/post-map-targets-json5-v-json-h_latencies.csv: 76.038368

[seancribbs]

What is meant by "mean 95th"? Is that actually the average of the 95th percentile? (Same goes for median.) Are there basho_bench graphs?

[seancribbs]

This whole case expression can be simplifed to:

is_datatype(BProps) orelse lists:member({consistent, true}, BProps)

[zeeshanlakhani]

derp on me... totally right.

[zeeshanlakhani]

@seancribbs was using the scripts that Ryan had been using in https://github.com/basho/yokozuna/tree/develop/misc/bench/bin (and documented in https://github.com/basho/yokozuna/blob/develop/docs/BENCHMARKING.md). And, yes, mean 95th is the avg of all 95th percentiles and median is the average of all the medians.

If you want I could showcase some graphs, but I thought to use what we had been doing previously.

[zeeshanlakhani]

@seancribbs new/different runs (same config), but showcasing the graphs.

develop

w/ perf update

[seancribbs]

👍 263c265

I'm convinced this is a significant-enough change. Good work!

[zeeshanlakhani]

@borshop merge