Bigchunks are held as a slice of sub-chunks, and a new sub-chunk is added every N samples by reference to remainingSamples which counts down from the maximum (currently 120).

remainingSamples is not initialized in Unmarshal(), so after an ingester hand-over a new sub-chunk is created immediately.

Leaving the smaller sub-chunk reduces overall compression ratios and slows everything down a bit.

While it would be expensive to count remainingSamples accurately, it seems to me we could estimate it based on the byte size of the last sub-chunk. We can count samples at the same time as we determine first and last sample for each sub-chunk.