Apache Drill already has a queueing feature based on ZK semaphores. We did a bit of testing to show that the feature does, in fact work. We propose to enhance the feature with some light revisions to make work with the "managed" external sort and the newly-added spilling feature for the hash agg operator. The key requirement is to build on what we have for now; we may want to tackle a larger project to create a more complete solution later.
- Two ZK-based queues called the “small” and “large” query queues.
- A threshold, call it T, given as a query cost, to determine the queue into which a query will go.
- Admit levels for the two queues: call them Qs and Ql.
Basically, when a query comes in:
- Plan the query as usual.
- Obtain the final query cost from the planner, call this C.
- If C<T, the query goes into the small queue, else it goes into the large queue.
- Suppose the small queue. Ask ZK if the query can run.
- ZK checks if Qs queries are already running. If so, the query waits, else the query runs.
The proposed changes include:
- Refactor the code to provide a queueing API that supports a variety of queuing mechanisms.
- Provide three: the null queue (default), an in-process queue (for testing) and the ZK queues.
- Modify the query profile web UI to show two new bits of information about queues:
- The queue to which the query was sent.
- The total planning cost.
- Modify the query profile web UI to show two memory assignment numbers:
- Total memory allocated to the query
- Memory per sort or hash-add operator
Then, add to the queue mechanism the ability to do memory assignment:
- Provide a weight, W: every small query gets 1 unit, every large query gets W units.
- Use the queue admit levels to determine total units: U = Qs + W * Ql.
- Obtain total direct memory from the system. M.
- Subtract a reserve percent R for overhead.
- Do the math to get the memory per query for each query:
- For the small queue: (M - R) / U
- For the large queue: (M - R) / U * W
- Use this memory amount as the “memory per query” number in the existing sort/hash-agg memory assignment (instead of the fixed 2 GB.)
The result will be a nice incremental addition to what we already have, and should make it a bit easier people to actually use the feature (because they can see the planning numbers and see the queues used, allowing them to effectively tune the system.)
The API used for the above features also allow third parties to add on a more robust admission control feature as needed, perhaps tying into an existing queueing mechanism of their choice.