Infrastructure » sepia

To reshard a bucket will block writes now, and the blocking time increases with objects in the bucket. This is not user-friendly, we should reduce blocking time as far as possible. Here is the design:

Split the bucket resharding into two phases: logrecord and progress. A log will be written with index operation to src shards in first phase, and the client writes will not be blocked; we then block the client writes, go through the recording log and copy the changed index entries to dest shards in second phase. In this way, we can greatly reduce the time blocking client writes in resharding a bucket.