Seal data blocks and close encoders when they are sufficiently old

[xichen2020]

cc @robskillington

When the blocks are old enough, there should be no writes going into those blocks, which means we can pull out the raw data out of the encoders, store them in the blocks, and close the encoders, which can save us the memory overhead of the encoder themselves.

[coveralls]

Coverage increased (+0.04%) to 76.488% when pulling f1efa19 on xichen-seal-blocks into ce94736 on master.

[robskillington]

if encoder := b.encoder; encoder != nil && b.writable { 
  b.ctx.RegisterCloser(encoder.Close)
} else if stream := b.stream; stream != nil && !b.writable {
  //...
}

Perhaps?

[xichen2020]

hm.I feel conditioning on b.writable is easier to comprehend than mixing it up with nil checks for encoders and streams

[robskillington]

Fair enough. You could do something that I've done in other places successfully to avoid this type of big branching:

cleanup := func() {
    b.ctx.Close()
    b.ctx = nil
    b.encoder = nil
    b.stream = nil
}
if b.writable {
    // If the block is not sealed, we need to close the encoder.
    if encoder := b.encoder; encoder != nil {
        b.ctx.RegisterCloser(encoder.Close)
    }
    cleanup()
    return
}
if stream := b.stream; stream != nil {
    b.ctx.RegisterCloser(func() {
        if bytesPool := b.opts.GetBytesPool(); bytesPool != nil {
            segment := stream.Segment()
            stream.Reset(m3db.Segment{})
            bytesPool.Put(segment.Head)
            bytesPool.Put(segment.Tail)
        }
        stream.Close()
    })
}
cleanup()

[xichen2020]

sure thing, updated.

[robskillington]

LGTM

[coveralls]

Coverage increased (+1.1%) to 77.578% when pulling 443c904 on xichen-seal-blocks into ce94736 on master.