DRILL-5512: Standardize error handling in ScanBatch#838
DRILL-5512: Standardize error handling in ScanBatch#838paul-rogers wants to merge 3 commits intoapache:masterfrom
Conversation
Standardizes error handling to throw a UserException. Prior code threw various exceptions, called the fail() method, or returned a variety of status codes.
1. Backed out UserError changes. Repurposed existing categories per Parth’s suggestions. 2. Revised ScanBatch to use the existing categories. The result is that this PR is now branched directly from master.
| */ | ||
| RESOURCE = 7; | ||
| /* equivalent to SQLNonTransientException. | ||
| * - out of memory |
There was a problem hiding this comment.
for out of memory we throw "memoryError" which internally set's the errorType as RESOURCE not SYSTEM.
| * | ||
| * <pre> | ||
| * equivalent to SQLNonTransientException. | ||
| * - out of memory |
|
+1 (pending minor comment). |
paul-rogers
left a comment
paul-rogers
left a comment
There was a problem hiding this comment.
Thanks for the review!
| * | ||
| * <pre> | ||
| * equivalent to SQLNonTransientException. | ||
| * - out of memory |
| */ | ||
| RESOURCE = 7; | ||
| /* equivalent to SQLNonTransientException. | ||
| * - out of memory |
| logger.debug("Caught Out of Memory Exception", e); | ||
| clearFieldVectorMap(); | ||
| return IterOutcome.OUT_OF_MEMORY; | ||
| throw UserException.memoryError(e).build(logger); |
There was a problem hiding this comment.
The non-managed external sort spills to disk in case it receives this outcome. I do not know if there are other operators that handle this outcome. Are all the pre-requisite changes (to handle this change) already committed?
There was a problem hiding this comment.
As it turns out, the idea of the OUT_OF_MEMORY return code works better in theory than in practice. No reader correctly handles this case. Let's say we have three columns (a, b, c). Let say that column c needs to double its vector, but hits OOM. No reader has the internal state needed to hold onto the value for c, unwind the call stack, then on the next next() call, rewind back to the point of writing c into the in-flight row.
Moving forward, we want to take a broader approach to memory: budget sufficient memory that readers can work. Modify the mutators so that they enforce batch size limits so that the reader operates within its budget.
As we move to that approach, the OUT_OF_MEMORY status will be retired.
The JIRA mentions another JIRA that holds a spec for all this stuff; something we discussed six months ago, but did not have time to implement then.
This all merits a complete discussion; maybe we can discuss the overall approach in that other JIRA.
There was a problem hiding this comment.
Makes sense.
Asking the question a different way.. To avoid regressions, should the pre-requisite changes (DRILL-5211 and pertinent tasks) be committed before this patch? Or since the readers do not correctly handle the case anyway, there will be no difference?
There was a problem hiding this comment.
Good question. Yes, since the FragmentExecutor already handles errors and unwinds, we just exploit that (existing, working) path instead of the (also existing, but harder-to-keep working) path of fail/STOP.
The (managed) external sort, for example, reports all its errors via exceptions; it does not use fail/STOP. The fragment executor recovers just fine.
There was a problem hiding this comment.
I am not sure if this specific line change is required, so please correct me if I am wrong. Thinking out loud..
There are three places in ScanBatch where OutOfMemoryException is handled. Since OutOfMemoryException is an unchecked exception, I could not quickly find all the calls which trigger the exception in this method.
The first case and second case are similar in that reader.allocate(...) fails. So although there is no unwind logic, seems to me, this case is correctly handled as no records have been read, and so there is no need to unwind. Say this triggers spilling in sort, then the query could complete successfully, if allocate succeeds next time (and so on). Am I following this logic correctly?
But this does not seems to be case, as TestOutOfMemoryOutcome triggers an OutOfMemoryException during "next" allocation, and all tests are expected to fail.
And then, there is the third case, which is a general catch (e.g.reader.next() throws OutOfMemoryException). And as you mentioned, readers cannot unwind, so that correctly fails the fragment.
There was a problem hiding this comment.
You raise two good points.
The first is that Drill tends to use unchecked exceptions because, well, they are convenient and Drill must handle such unchecked exceptions as NPE, illegal state, illegal argument and so on.
But, normal Java practice is to declare exceptions so that callers know what they should handle. See DRILL-5386 asks to rationalize exception use. Comments in that ticket from project veterans shows some degree of resistance to the idea of checked exceptions. So, yes we must expect any unchecked exception from any method. This is why operators should handle all exceptions, and why we need code to sort out exceptions based on type.
The analysis of OOM is correct, but omits context. It is seldom (never?) that a sort sits directly above a scan. Seems most often there is a filter or project between them. If the scan hits OOM, it is not likely because it has exhausted the memory limit on the scan operator: each operator defaults to 10 GB limit. Instead, it is likely that overall system memory is exhausted. So, what is likely to happen? Each operator between the scan and sort must handle the OUT_OF_MEMORY status by bubbling it up the call stack. Let's assume that works.
The sort now wants to spill. Spilling is an activity that requires memory to perform. Spilling requires a merge phase to combine the records from buffered batches in sort order so that the spilled run is sorted. That is, the sort must allocate a batch, often many MB in size. (Spilled runs must be sizable so we can limit the number of spilled runs merged in the final merge phase.)
So, the sort tries to allocate vectors for the merge batch and... The allocation fails. Why? Because we are out of system memory -- that's why the scanner triggered an OOM.
I can find no code that sets up this out-of-system-memory condition to verify that existing code works. I think we were taking it on faith that this behavior actually works.
Moving forward, we are working towards a workable solution. Assign the scan some amount of memory, and limit batches to fit within that memory. Give the sort a certain amount of memory, and have it manage within that memory so that when a spill occurs, the sort has sufficient memory to create the required merge batches as part of the spill.
Finally, let's consider the case you outlined: the scan fails with OOM on the initial allocation. The initial allocation is often small; the scan goes through many vector doublings to read the full complement of rows. (At least, thats what I've observed empirically; perhaps the original intent was different.) Let's say we tried to recover from an initial allocation failure.
Say we have a row with five columns. We allocate three, but fail on the fourth. Say the fourth is a Varchar: has two vectors: offset and data. The current logic releases the partially-allocated vectors, which is good. OUT_OF_MEMORY is returned and the vectors are reallocated if memory could be released. Sounds good.
But, most queries run in multiple threads. If one hits OOM, then the others probably will as well. The actual system memory is a shared resource, but there is no coordination. A scan might release its partially-allocated vectors so the sort can, in theory, spill. But, that memory might immediately be grabbed by some other thread, resulting in a sort spill OOM. In practice, however, the initial vector allocations are much smaller than the merge batch, so it is only slightly useful to free up the initial allocation. That initial allocation, plus luck that some other thread has freed enough memory, might allow us to spill. But it is a crap-shoot.
In short, even if this logic might possibly work in some scenarios in a single-threaded query, it is too chaotic to work in general with many threads. And, of course no tests exist for either case so we are just guessing.
All-in-all, the above argues strongly that the path forward is to:
- Rationalize error handling: OOM errors cause query failure.
- Design a memory assignment system so that operators live within a budget.
- Design tests to ensure that the system works rather than relying on hope.
This checkin is a step toward goal 1. The external sort revision, hash agg spilling and other projects are steps toward goal 2. We continue to chip away at our ability to do goal 3.
Given all of this, can you suggest how we could gather evidence that the current OUT_OF_MEMORY status is actually working in any actual queries? Or, do we have a tough case of comparing concrete changes against an aspiration for how the system work might?
More practically, with the change, OOM will fail the query. Previously, there is some chance that Drill might recover from an OOM. But, we have no tests and no statistics. Is it worth risking that hypothetical for a concrete step in the right direction. I don't think we have the answer and that, itself, is a problem. Thoughts?
There was a problem hiding this comment.
I agree. I was hoping there are tests to show current handling of OUT_OF_MEMORY as well.
There was a problem hiding this comment.
As it turns out, in addition to fixing ScanBatch, there is a need for a parallel implementation to support a the size-aware vector "writer". That new version was modified to allow extensive unit testing of all error conditions. Can't do that as easily with ScanBatch itself because it has messy references to other parts of Drill, which have references, which have references... Pretty soon all of Drill is needed to do the tests, which means using the current injection framework (which we could do.)
| } | ||
| addImplicitVectors(); | ||
| } catch (ExecutionSetupException e) { | ||
| this.context.fail(e); |
There was a problem hiding this comment.
This call triggers query failure (stopping the fragment, notifying the Foreman, and cancelling other fragments, etc.). What is the flow after this change?
Similar changes below.
There was a problem hiding this comment.
Throwing an exception, it turns out, does exactly the same: it cancels the query and causes the fragment executor to cascade close() calls to all the operators (record batches) in the fragment tree. It seems some code kills the query by throwing an exception, other code calls the fail method and bubbles up STOP. But, since the proper way to handle STOP is to unwind the stack, STOP is equivalent to throwing an exception.
The idea is, rather than have two ways to clean up, let's standardize on one. Since we must handle unchecked exceptions in any case, the exception-based solution is the logical choice for standardization.
|
+1 |
Standardizes error handling to throw a UserException. Prior code threw various exceptions, called the fail() method, or returned a variety of status codes. closes apache#838
3 participants
Standardizes error handling to throw a UserException. Prior code threw
various exceptions, called the fail() method, or returned a variety of
status codes.
Note: this PR depends upon DRILL-5511, PR #836. For this PR, please review only the changes introduced in DRILL-5512. Commit this after committing DRILL-5511.