For some workloads, it may be difficult to predict in advance how large the data set will grow. Small data sets are handled efficiently by btree, out-of-cache data sets are more efficient with LSM.

Since WiredTiger uses the btree format for each chunk of an LSM tree, we could start by creating a btree, then once some threshold size is reached, switch to LSM. Applications could specify "type=auto" or similar when creating such a table.

The transition from btree to LSM would involve existing cursors noticing the change and reconfiguring, so it would require a step like __clsm_enter even before the transition.