Proposal: Branch Read Optimization

[N-o-Z]

Proposal for branch read optimization

[itaiad200]

Too late to update it now. There's already an entry in staging. If another read request comes, it should be able to see it.

[N-o-Z]

You are absolutely right - we must first set dirty bit before any write

[ozkatz]

Out of curiosity - do we know how common this case is? (i.e. % of reads that go to "clean" branches?) Might help us understand if the added complexity is worth it.

Perhaps a more granular approach could be beneficial: we can go further than a single boolean and maintain a small bitmap: one bit per range - essentially lakeFS' take on dirty pages.
Instead of reading the bool, read the bitmap; and if the requested range(s) are 0, you're good to read just the underlying committed data. if it's 1, check staging. on write, mark 1 for any range affected by the write.