ClickHouse: Allow cancelling queries in the middle of blocking operations
Queries can be canceled only at specific points between blocks processing. However, a query may get stuck on a blocking operation like reading/writing from/to a socket. In this case, it has to wait for receive_timeout/send_timeout (default is 300 seconds, can be changed to a higher value) and cannot be killed/canceled.
However, we know ids of all threads executing a query. So we can send a signal to these threads and make them exit from a blocking syscall with EINTR. After that, we can check a flag in ThreadStatus/ThreadGroup before resuming the syscall and throw an exception if the query was canceled.
About this issue
- Original URL
- State: open
- Created a year ago
- Reactions: 7
- Comments: 15 (15 by maintainers)
@serxa, throwing inside the EINTR loop + cancellable mutex could suffice.
It makes sense, but it will not help when execution time is not limited. This issue is about
KILL QUERYandCtrl-C, not about max execution time.We can start by simply adding a check for a cancellation flag because we already have the (real-time) query profiler that sends a signal each 1 second by default.