Restore's per-chunk loop called blobCache.Get(blobHash) and sliced the
returned []byte to extract the chunk it actually wanted. Get reads the
entire blob from disk into memory — so for a 10 GB blob, every chunk
extraction was a 10 GB ReadFile to get back a few KB. With ~40k files
needing ~600ms per cache hit, that alone was burning ~6 hours of
wall-clock on a real restore.
Hot loop now:
- If the blob isn't cached: download (full plaintext into memory),
Put to disk cache, satisfy this chunk from the in-memory buffer.
- If it's cached: blobCache.ReadAt(hash, offset, length) — reads
only the chunk's bytes from the on-disk blob file.
ReadAt was already implemented on blobDiskCache; restore just wasn't
using it.
Debug timings from the user's photo-catalog restore showed
ms_cache_gets dominating every cache-hit file at 500-1000ms. With
ReadAt those should drop to sub-millisecond and the visible throughput
should be bound by single-stream blob download + decrypt/decompress
rather than disk-read amplification.
sneak
8036d93914