> Is the recent begin-concurrent branch get extra 40% TPS compare to 2017-07-31 version ? I doubt it. The server-process-edition branch uses latches to manage readers and writers, whereas begin-concurrent is based on snapshot isolation. This means that if you tweak the locking primitives in the VFS and combine it with something like the wal2 branch, begin-concurrent scales to hundreds of concurrent readers, whereas the locking required of readers by server-process-edition means that performance of both readers and writers starts to drop off after a few dozen reader threads are added. I suppose most users wouldn't really care - hardware that can run that many threads is quite uncommon. It is the reason begin-concurrent is preferred though.