feat(dropPrefix): add DropPrefixNonBlocking API

[NamanJain8]

Related to DGRAPH-3319

This PR adds DropPrefixNonBlocking API that can be used to logically delete the data for specified prefixes at a given timestamp. This also adds an equivalent API of older DropPrefix == DropPrefixBlocking.
DropPrefix now makes decision based on badger option AllowStopTheWorld whose default is to use DropPrefixBlocking.
With DropPrefixNonBlocking , the data would not be cleared from the LSM tree immediately. It would be deleted eventually through compactions.
This operation is useful when we don't want to block writes while we delete the prefixes (DropPrefix blocks writes).
It does this in the following way:

• Stream the given prefixes.
• Write them to skiplist and handover that skiplist to DB.

This change is 

[manishrjain]

Approved, but address the comments carefully.

Reviewed 1 of 1 files at r2, 1 of 1 files at r3, 1 of 1 files at r4.
Reviewable status: all files reviewed, 12 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @NamanJain8)

---

db.go, line 1879 at r4 (raw file):

				list.Kv = append(list.Kv, kv)

				if db.opt.NumVersionsToKeep == 1 {

If you just have a delete marker at higher timestamp, everything below it would be considered deleted anyway.

For NumVersionsToKeep > 1 (in Dgraph), you'd continue to iterate upon it and generate more delete markers. You just need to generate one per key.

---

db.go, line 1883 at r4 (raw file):

				}

				if item.DiscardEarlierVersions() {

Don't care about this. Both of these if conds are not required.

---

db.go, line 1890 at r4 (raw file):

		}

		var wg sync.WaitGroup

move this outside.

---

db.go, line 1894 at r4 (raw file):

		initSize := int64(float64(db.opt.MemTableSize) * 1.1)

		handover := func(force bool) error {

if cbuf.Len > whatever, then sort, and create a skiplist, handover.

---

db.go, line 1897 at r4 (raw file):

			for id, b := range builderMap {
				sl := b.Skiplist()
				if force || sl.MemSize() > db.opt.MemTableSize {

don't need to check this. Just use cbuf size.

---

db.go, line 1908 at r4 (raw file):

			return nil
		}

var cbuf *z.Buffer

---

db.go, line 1910 at r4 (raw file):

		stream.Send = func(buf *z.Buffer) error {
			err := buf.SliceIterate(func(s []byte) error {

You have a cbuf outside, and you append buf to cbuf. And then sort cbuf by keys. Create a skiplist out of it using builder, and then handover.

---

db.go, line 1918 at r4 (raw file):

					builderMap[kv.StreamId] = skl.NewBuilder(initSize)
				}
				builderMap[kv.StreamId].Add(y.KeyWithTs(kv.Key, ts), y.ValueStruct{Meta: bitDelete})

this would be expensive. You'd rather use skiplist builder. And for that, use the idea above.

---

db.go, line 1942 at r4 (raw file):

			return errors.Wrapf(err, "While dropping prefix: %#x", prefix)
		}
	}

wg.Wait() here.

[manishrjain]

let's review this after changes.

Reviewable status: all files reviewed, 13 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @NamanJain8)

---

db.go, line 1843 at r1 (raw file):

Previously, NamanJain8 (Naman Jain) wrote…

Done.

This should be done via an option. DropPrefix should work in either blocking or non-blocking mode.

You could still have these public funcs, but DropPrefix should choose one based on Badger.Options.

DropPrefixNonBlocking and DropPrefixBlocking.

---

db.go, line 1877 at r4 (raw file):

				kv := y.NewKV(a)
				kv.Key = ka
				list.Kv = append(list.Kv, kv)

You know the version to use here.

[NamanJain8]

Dismissed @jarifibrahim and @manishrjain from 9 discussions.
Reviewable status: 1 of 4 files reviewed, 4 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @manishrjain)

---

db.go, line 1843 at r1 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

This should be done via an option. DropPrefix should work in either blocking or non-blocking mode.

You could still have these public funcs, but DropPrefix should choose one based on Badger.Options.

DropPrefixNonBlocking and DropPrefixBlocking.

Done. We have BlockWritesOnDrop DB option.

---

db.go, line 1845 at r1 (raw file):

Previously, jarifibrahim (Ibrahim Jarif) wrote…

return error

Done.

---

db.go, line 1877 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

You know the version to use here.

Done.

---

db.go, line 1879 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

If you just have a delete marker at higher timestamp, everything below it would be considered deleted anyway.

For NumVersionsToKeep > 1 (in Dgraph), you'd continue to iterate upon it and generate more delete markers. You just need to generate one per key.

Done.

---

db.go, line 1883 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

Don't care about this. Both of these if conds are not required.

Done.

---

db.go, line 1890 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

move this outside.

Done.

---

db.go, line 1894 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

if cbuf.Len > whatever, then sort, and create a skiplist, handover.

Done.

---

db.go, line 1897 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

don't need to check this. Just use cbuf size.

Done.

---

db.go, line 1908 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

var cbuf *z.Buffer

Done.

---

db.go, line 1910 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

You have a cbuf outside, and you append buf to cbuf. And then sort cbuf by keys. Create a skiplist out of it using builder, and then handover.

Done.

---

db.go, line 1918 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

this would be expensive. You'd rather use skiplist builder. And for that, use the idea above.

Done

---

db.go, line 1942 at r4 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

wg.Wait() here.

Done.

[manishrjain]

Reviewed 3 of 3 files at r5.
Reviewable status: all files reviewed, 10 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @NamanJain8)

---

db.go, line 1864 at r5 (raw file):

		}

		var kvs []*pb.KV

Anytime Go memory gets involved, it kills it.

---

db.go, line 1865 at r5 (raw file):

		var kvs []*pb.KV
		err := cbuf.SliceIterate(func(slice []byte) error {

cbuf.SliceSort.

---

db.go, line 1884 at r5 (raw file):

			return bytes.Compare(kvs[i].Key, kvs[j].Key) < 0
		})
		for _, kv := range kvs {

iterate over cbuf and do the generation of the skiplist.

---

db.go, line 1885 at r5 (raw file):

		})
		for _, kv := range kvs {
			b.Add(y.KeyWithTs(kv.Key, kv.Version+1), y.ValueStruct{Meta: bitDelete})

Just use the same Version.

---

db.go, line 1917 at r5 (raw file):

			kv.Key = ka
			kv.Version = item.Version()
			list.Kv = append(list.Kv, kv)

You could just generate a key with ts.

---

db.go, line 1924 at r5 (raw file):

		stream.Send = func(buf *z.Buffer) error {
			sz := buf.LenNoPadding()
			dst := cbuf.Allocate(sz)

you could iterate over buf, and do slice allocates over cbuf to write the keywithTs ONLY.

---

db.go, line 1925 at r5 (raw file):

			sz := buf.LenNoPadding()
			dst := cbuf.Allocate(sz)
			y.AssertTrue(sz == copy(dst, buf.Bytes()))

Iterate over buf here, and use the same KV.

---

options.go, line 690 at r5 (raw file):

//
// The default value of BlockWritesOnDrop is true.
func (opt Options) WithBlockWritesOnDrop(b bool) Options {

WithBlockingDrops(b bool)

[manishrjain]

Reviewable status: all files reviewed, 10 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @NamanJain8)

---

options.go, line 690 at r5 (raw file):

Previously, manishrjain (Manish R Jain) wrote…

WithBlockingDrops(b bool)

A better name would be -- AllowStopTheWorld -- true by default. And false for Dgraph.

[manishrjain]

Approved. Nice work!

Reviewed 5 of 5 files at r6.
Reviewable status: all files reviewed, 10 unresolved discussions (waiting on @ahsanbarkati, @jarifibrahim, and @NamanJain8)