Always set BlockSize in encoder.

[martinmr]

Internal Ref:

• DGRAPH-1267

---

Rollup was setting the BlockSize of the encoder for multi-part list but not for
normal lists.

This caused ApproxLen to return 0. Then the intersect algorithm exited early in this case.
Existing data will be encoded with the right block size the next time the list is rolled up.

This PR does the following.

1. Set the blocksize.
2. Add an assert in rollup to verify that the blocksize is always the right value.
3. Eliminate the check in the intersect algorithm to not return immediately even if ApproxLen
   returns 0. There's no difference if the actual length is zero since the smallest list is picked
   to perform the iteration. If the list has data but a block size of zero, removing this check allows
   the query to return the right result. Added a test to verify this scenario.
4. Added missing tests for the binary search intersect algorithm.

Fixes #5102

---

This change is 

Docs Preview:

[danielmai]

Does this affect only trigram indices or @filter (intersections) with other indices like eq() too?

Reviewable status: 0 of 1 files reviewed, all discussions resolved (waiting on @manishrjain)

[martinmr]

It affects all the queries that:

1. Use the Uids method
2. The list only has an immutable layer.
3. The list of uids needs to be intersected.

I'll update the description.

Reviewable status: 0 of 1 files reviewed, all discussions resolved (waiting on @manishrjain)

[manishrjain]

Reviewable status: 0 of 1 files reviewed, 1 unresolved discussion (waiting on @danielmai, @manishrjain, and @martinmr)

---

posting/list.go, line 985 at r1 (raw file):

	res := make([]uint64, 0, len(l.mutationMap)+codec.ApproxLen(l.plist.Pack))
	out := &pb.List{}
	if len(l.mutationMap) == 0 && opt.Intersect != nil && len(l.plist.Splits) == 0 {

Removing this would affect performance quite a bit I think. Better to determine why IntersectCompressedWith fails.

[manishrjain]

Got a few comments. Please address before merging.

Reviewable status: 0 of 4 files reviewed, 4 unresolved discussions (waiting on @danielmai, @manishrjain, and @martinmr)

---

algo/uidlist.go, line 99 at r2 (raw file):

	lq := len(q)

	if ld == 0 || lq == 0 {

We could perhaps keep lq at least -- that is exact and not approximate.

---

posting/list.go, line 935 at r2 (raw file):

	})
	// Finish  writing the last part of the list (or the whole list if not a multi-part list).
	x.Check(err)

This can perhaps also be an error return.

---

posting/list.go, line 938 at r2 (raw file):

	plist.Pack = enc.Done()
	if plist.Pack != nil {
		x.AssertTrue(plist.Pack.BlockSize == uint32(blockSize))

If there's an error return by this func, please return error. I think we should go away from doing Asserts now.

[martinmr]

Reviewable status: 0 of 4 files reviewed, 4 unresolved discussions (waiting on @danielmai and @manishrjain)

---

algo/uidlist.go, line 99 at r2 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

We could perhaps keep lq at least -- that is exact and not approximate.

Done.

---

posting/list.go, line 985 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Removing this would affect performance quite a bit I think. Better to determine why IntersectCompressedWith fails.

Done. Not removing this anymore.

---

posting/list.go, line 935 at r2 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

This can perhaps also be an error return.

Done.

---

posting/list.go, line 938 at r2 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

If there's an error return by this func, please return error. I think we should go away from doing Asserts now.

Done.