← Back to team overview

randgen team mailing list archive

Solution for the QueryTimeout + Transform interaction



It turns out that if you do experiments with SLEEP() or BENCHMARK(), you get weird behavior since such a query, even if killed, will return a result set and the server will terminate the connection after the result has been returned. I assume this is by design because no rows are processed by those functions and therefore neither KILL only works after the fact -- usually it is a check if you have been killed is made as each row is processed, and here we are not processing any rows.

The Executor is not set up to deal with this situation, where you both get a result set on the query but then "connection lost" on the next query. I do not think there is a need to make it even more complicated in order to account for this very rare scenario.

So, to do experiments, use something like SELECT SUM(RAND()) FROM C,CC,D,E to force actual rows to be processed.

So, in order for me to fix the issue, all I had to do was fortify some Transforms with extra DROP TABLE IF EXISTS and then add the various "connection lost" error codes to the mysql_grouping_errors table in lib/GenTest/Transform.pm (the name is rather misleading). Then, the existing mechanism will make sure the test continues properly.

I have not implemented the separation of STATUS_QUERY_KILLED from STATUS_SERVER_CRASHED that we discussed, since it was not necessary to fix this particular issue. Some day, it would be a nice thing to have.

Philip Stoev