It is possible to use data from multiple Prometheus instances without manually rewriting the database: simply copy blocks from one Prometheus data directory to the target Prometheus data directory.

For example, Prometheus data directories usually look like this:

01HQAK3HN2YAJ8XJPB3B5P88XX/
01HQBW9EBQEANHPGTEP5YDF5D0/ 
01HQC355KQ5V8WPCFKA3DJ8E9H/
chunks_head/
queries.active 
wal/

In this case the first 3 subdirectories starting with 01HQ* are blocks that have been compacted.

The main caveat is this works well enough for already compacted TSDB blocks, but it doesn't necessarily work well for the head block (the block that holds the most recent data), and so there is a chance we are losing the last ~2 hours of the Prometheus databases we are merging.

It would be ideal if we can simply flush the head block with the Prometheus API or with promtool. There are a number of proposals, but nothing that looks like it has made it to Prometheus yet.

Alternatives I'm thinking of are playing with snapshots to force writing the head block, or forcing it directly with a custom tool that calls FlushWAL() on the Prometheus database. While not perfect, this approach to combining multiple Prometheus databases is more promising than other options.

This issue is meant to keep track of progress and findings related to this topic.