hurikhan's suggestion makes good sense.

Had planned to release LMDB 1.0 this summer, but I've been a bit too busy with other things to get all the feature branches merged. Sadly, I would advise you not to hold your breath.

In T9805#181577, @bruns wrote:

In T9805#181576, @hyc wrote:

The MDB_val for the data contains a size. Divide by the size of a single DUPFIXED item to get the number of items returned in that page.

Yes, makes sense. For put, the second MDB_val is required as the fixed val size is not known to LMDB until the first put?

In T9805#181560, @bruns wrote:

In T9805#181517, @hyc wrote:

Fwiw, @poboiko is correct here.

In T9805#181157, @poboiko wrote:

Just because you hide the RMW, it does not mean it does not happen. For LMDB, duplicate keys are just a plain array, see e.g. MDB_APPENDDUP in http://www.lmdb.tech/doc/group__mdb.html#ga4fa8573d9236d54687c61827ebf8cac0

Nothing on that link implies use of an array anywhere.

MDB_GET_MULTIPLE returns "up to a page of duplicate data items from current cursor position", so apparently the values for DUPFIXED are packed, or is the page created during the mdb_cursor_get call?

Fwiw, @poboiko is correct here.