I've been stepping through some decompression code to try to figure out why zstd memory use is much higher than other compression types.

I see that Reset() uses the buffer d.current.b:

Then in my case DecodeAll runs and the buffer grows to ~300KB.

Each time Read() is called, it advances the start of the buffer:

After we read the whole thing, b is still pointing into a 300KB buffer, but towards the end.
When we re-use the Decoder, this buffer has len zero and cap ~35KB. Because 300KB is needed, a new block is allocated.

Could we have the Decoder hold a reference to the beginning of the buffer, so the whole size can be re-used?

Apologies if I'm mis-reading this.