Issue #2849
It turns out that the lion's share of the time was consumed
by byte-by-byte operations. There were two factors:
1. Calling panic in a function somehow makes it more expensive.
2. Calling a function through an interface also adds up.
This PR makes spot improvements to address two cases:
1. The tokenizer was calling bytes.Buffer.WriteByte for every byte
of the input query.
2. The EncodeSQL function in sqltypes.Value was calling
a wrapper (that had a panic), which in turn called
buffer.WriteByte.
I've now written a bytes2.Buffer package that implements a subset
of bytes.Buffer, but more efficiently.
Resulting improvements:
VTGate New:
BenchmarkWithNormalizer-4 300 5694660 ns/op
VTGate Old:
BenchmarkWithNormalizer-4 100 10927765 ns/op
VTTablet New:
BenchmarkExecuteVarBinary-4 300 4751125 ns/op
VTTablet Old:
BenchmarkExecuteVarBinary-4 100 14888528 ns/op
These numbers can further be improved as mentioned in the issue.
But the work to get to the next level is more involved.
xichengliudui
Sugu Sougoumarane
Alain Jobart
Anthony Yeh
Sugu Sougoumarane
Ryszard Szopa