fix(rollups): rollup a batch if more than 2 seconds elapsed since last batch

[parasssh]

fix(rollup): rollup a batch every 2 seconds or 64 keys

---

This change is 

[martinmr]

Reviewable status: 0 of 2 files reviewed, all discussions resolved (waiting on @manishrjain and @martinmr)

[danielmai]

What's the benchmark result? Can you paste the results here?

Reviewable status: 0 of 2 files reviewed, all discussions resolved (waiting on @manishrjain)

[parasssh]

What's the benchmark result? Can you paste the results here?

Reviewable status: 0 of 2 files reviewed, all discussions resolved (waiting on @manishrjain)

1. Benchmark with using atomic operations on the lastSentTs

Running tool: /usr/local/go/bin/go test -benchmem -run=^$ github.com/dgraph-io/dgraph/posting -bench ^(BenchmarkIncrRollupAtomic)$

[Decoder]: Using assembly version of decoder
badger 2020/07/30 19:27:26 INFO: All 0 tables opened in 0s
goos: linux
goarch: amd64
pkg: github.com/dgraph-io/dgraph/posting
BenchmarkIncrRollupAtomic-8   	18172532	        65.5 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	github.com/dgraph-io/dgraph/posting	1.277s

2. Benchmark without using atomic on lastSentTs.

Running tool: /usr/local/go/bin/go test -benchmem -run=^$ github.com/dgraph-io/dgraph/posting -bench ^(BenchmarkIncrRollupAtomic)$

[Decoder]: Using assembly version of decoder
badger 2020/07/30 19:31:23 INFO: All 0 tables opened in 0s
goos: linux
goarch: amd64
pkg: github.com/dgraph-io/dgraph/posting
BenchmarkIncrRollupAtomic-8   	17659142	        65.2 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	github.com/dgraph-io/dgraph/posting	1.241s

So, (1) and (2) are comparable at ~65 ns/op

However, I have not done the benchmark directly on master as it is today. Let me do that next.

[parasssh]

3. Benchmark on master without any changes.

Running tool: /usr/local/go/bin/go test -benchmem -run=^$ github.com/dgraph-io/dgraph/posting -bench ^(BenchmarkIncrRollupAtomic)$

[Decoder]: Using assembly version of decoder
badger 2020/07/30 19:35:36 INFO: All 0 tables opened in 0s
goos: linux
goarch: amd64
pkg: github.com/dgraph-io/dgraph/posting
BenchmarkIncrRollupAtomic-8   	60290652	        18.8 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	github.com/dgraph-io/dgraph/posting	1.177s

So, on master, it is ~19 ns/op but it suffers from the issue that rollups dont happen as often.
With my branch it is ~65 ns/op as seen in the previous comment but it fixes above issue by forcing rollup if >2 sec has elapsed since the last time.

FWIW, I live-loaded 21 million data set with master and my branch. It took 13m30sec and 13m49sec respectively. So, not much change there.

cc @manishrjain

[parasssh]

The time.Now().Unix() is the culprit. It is expensive in itself. It adds ~40ns/op.

[parasssh]

Based on the above comments, I think we can still use this fix.
The Benchmark is a tight loop that does millions of operations in 1-2 sec. This may not be the real world scenario.

[martinmr]

Reviewed 1 of 2 files at r2.
Reviewable status: all files reviewed, 1 unresolved discussion (waiting on @manishrjain and @parasssh)

---

posting/mvcc.go, line 120 at r2 (raw file):

dorollup

nit: doRollup