Write-back techniques at datastore accelerators

A fleet of query accelerator nodes is established for a data store. A particular node determines a write throughput limit for back-end writes from the fleet to the data store. A pool of write-back threads is established at the node. When a write request is received at the node, an entry is added to a selected partition of a write-back journal, and a write completion response is provided to the requester without applying the write to the data store. A write-back thread from the pool eventually schedules the back-end write.