The heretic's approach:
- have each thread write INSERT INTO <table> VALUES (<values>); statements into a text file
- when triggered (and/or periodically), each thread closes and renames the text file, then opens a new file
- main thread processes the renamed text files into the DB and renames them again (or deletes them if you don't require rollforward capability)
No pesky synchronization of databases, predictable effort storing the metrics, just one connection doing all the writes in a batch transaction.
Alternatively, instead of full statements, have the metric threads write csv files containing just the data and import those either directly in the shell or via a CSV table valued function.
INSERT INTO <table> SELECT * FROM CSV('<filename>');