bnclabs / robt

Read only BTree, immutable index for disk/SSD.

Index key,value dataset using a Read Only Btree. Refer to rustdoc link (above) for details.

Contribution

• Simple workflow. Fork - Modify - Pull request.
• Before creating a PR,
  • Run make build to confirm all versions of build is passing with 0 warnings and 0 errors.
  • Run check.sh with 0 warnings, 0 errors and all testcases passing.
  • Run perf.sh with 0 warnings, 0 errors and all testcases passing.
  • Install and run cargo spellcheck to remove common spelling mistakes.
• Developer certificate of origin is preferred.

robt's Issues

Caching btree-blocks.

One of the advantage of using read-only immutable btree index is that caching the btree-blocks are very straight-forward. A typical requirement would be:

• Have an in-memory (key,value) map where key is file-position as u64 type and value is Vec<Entry<K,V,D>>. Note that we can as well cache parsed blocks.
• We cannot cache all blocks, to limit caching we do the following.
  • Cache all inter-mediate blocks upto N levels.
  • Cache all intermediate blocks, but not leaf-blocks.
• Now caching leaf-blocks would require an eviction strategy. We can handle that as a separate issue.

Bloom filter for key lookup.

Get() operation into the index can be made faster by attaching a bloom-filter to the index. Typically it is expected to give false-positive by no false-negative. Bitmap index (implementing the Bloom) is already integrated with the index Builder type. We need to integrate them on the read-side. Specifically on the get() side.