feat(Dgraph): Add experimental cache for posting lists

[martinmr]

Add a cache for posting lists. There are still some issues (i.e failing Jepsen tests) so this feature
is being controlled by a flag for now and disabled by default.

---

This change is 

Docs Preview:

[manishrjain]

Get it reviewed by @balajijinnah as well. Looks alright to me, have a few comments.

Reviewed 8 of 8 files at r1.
Reviewable status: all files reviewed, 4 unresolved discussions (waiting on @martinmr and @vvbalaji-dgraph)

---

dgraph/cmd/alpha/run.go, line 203 at r1 (raw file):

	// Flag to enable posting list Ristretto cache.
	flag.Bool("pl_cache", false, "EXPERIMENTAL. Set to true to enable the posting list cache. "+

pl_cachemb

---

posting/lists.go, line 168 at r1 (raw file):

		for range ticker.C {
			ostats.Record(context.Background(),
				x.CacheInUse.M(int64(m.CostAdded()-m.CostEvicted())),

This would be tricky because we'll have many caches -- also from Badger.

---

posting/mvcc.go, line 268 at r1 (raw file):

// RemoveCacheFor will delete the list corresponding to the given key.
func RemoveCacheFor(key []byte) {
	lCache.Del(key)

Maybe not Del, instead do Set.

---

posting/mvcc.go, line 433 at r1 (raw file):

		return l, err
	}
	lCache.Set(key, l, 0)

The cost should never be zero.

[martinmr]

Reviewable status: all files reviewed, 4 unresolved discussions (waiting on @balajijinnah, @manishrjain, and @vvbalaji-dgraph)

---

dgraph/cmd/alpha/run.go, line 203 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

pl_cachemb

Done. If I am understanding this correctly, adding "mb" means we should pass the size of the cache in this flag and disable the cache if the value is zero, right?

---

posting/lists.go, line 168 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

This would be tricky because we'll have many caches -- also from Badger.

Ok. I have removed the metrics for now and added a log instead that prints the metrics once every minute.

---

posting/mvcc.go, line 268 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Maybe not Del, instead do Set.

Not sure if this is correct. What if the key is evicted, then we wouldn't update immediately.
@balajijinnah, what do you think?

---

posting/mvcc.go, line 433 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

The cost should never be zero.

If it's zero, Ristretto will use the default cost function.

[manishrjain]

Added some comments, please address.

Reviewed 7 of 7 files at r2.
Reviewable status: all files reviewed, 5 unresolved discussions (waiting on @martinmr and @vvbalaji-dgraph)

---

dgraph/cmd/alpha/run.go, line 203 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

pl_cachemb

pl_cache_mb

---

posting/lists.go, line 168 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

This would be tricky because we'll have many caches -- also from Badger.

Perhaps move to V(2) for now. @jarifibrahim should move all of these metrics out to prometheus metrics.

---

posting/lists.go, line 165 at r2 (raw file):

	go func() {
		m := lCache.Metrics
		ticker := time.NewTicker(60 * time.Second)

5 * time.Minute

---

posting/mvcc.go, line 268 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Maybe not Del, instead do Set.

Maybe add a note that we should perhaps try using Set with nil value to deal with jepsen failures.

---

posting/mvcc.go, line 282 at r2 (raw file):

func WaitForCache() {
	lCache.Wait()

Maybe comment this out for now. Add a note to investigate and better understand why Jepsen fails with cache on.

[martinmr]

Reviewable status: all files reviewed, 5 unresolved discussions (waiting on @balajijinnah, @jarifibrahim, @manishrjain, and @vvbalaji-dgraph)

---

dgraph/cmd/alpha/run.go, line 203 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

pl_cache_mb

Done.

---

posting/lists.go, line 168 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Perhaps move to V(2) for now. @jarifibrahim should move all of these metrics out to prometheus metrics.

Done.

---

posting/lists.go, line 165 at r2 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

5 * time.Minute

Done.

---

posting/mvcc.go, line 268 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Maybe add a note that we should perhaps try using Set with nil value to deal with jepsen failures.

Done.

---

posting/mvcc.go, line 282 at r2 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Maybe comment this out for now. Add a note to investigate and better understand why Jepsen fails with cache on.

Done.