Examples with AbstractBulkByScrollRequestBuilder org.elasticsearch.action.bulk.byscroll.AbstractBulkByScrollRequestBuilder used on opensource projects

Example 1 with AbstractBulkByScrollRequestBuilder

use of org.elasticsearch.action.bulk.byscroll.AbstractBulkByScrollRequestBuilder in project elasticsearch by elastic.

the class RethrottleTests method testCase.

private void testCase(AbstractBulkByScrollRequestBuilder<?, ?> request, String actionName) throws Exception {
    logger.info("Starting test for [{}] with [{}] slices", actionName, request.request().getSlices());
    /* Add ten documents per slice so most slices will have many documents to process, having to go to multiple batches.
         * we can't rely on all of them doing so, but 
         */
    List<IndexRequestBuilder> docs = new ArrayList<>();
    for (int i = 0; i < request.request().getSlices() * 10; i++) {
        docs.add(client().prepareIndex("test", "test", Integer.toString(i)).setSource("foo", "bar"));
    }
    indexRandom(true, docs);
    // Start a request that will never finish unless we rethrottle it
    // Throttle "forever"
    request.setRequestsPerSecond(.000001f);
    // Make sure we use multiple batches
    request.source().setSize(1);
    ListenableActionFuture<? extends BulkByScrollResponse> responseListener = request.execute();
    TaskGroup taskGroupToRethrottle = findTaskToRethrottle(actionName, request.request().getSlices());
    TaskId taskToRethrottle = taskGroupToRethrottle.getTaskInfo().getTaskId();
    if (request.request().getSlices() == 1) {
        assertThat(taskGroupToRethrottle.getChildTasks(), empty());
    } else {
        // There should be a sane number of child tasks running
        assertThat(taskGroupToRethrottle.getChildTasks(), hasSize(allOf(greaterThanOrEqualTo(1), lessThanOrEqualTo(request.request().getSlices()))));
        // Wait for all of the sub tasks to start (or finish, some might finish early, all that matters is that not all do)
        assertBusy(() -> {
            BulkByScrollTask.Status parent = (BulkByScrollTask.Status) client().admin().cluster().prepareGetTask(taskToRethrottle).get().getTask().getTask().getStatus();
            long finishedSubTasks = parent.getSliceStatuses().stream().filter(s -> s != null).count();
            ListTasksResponse list = client().admin().cluster().prepareListTasks().setParentTaskId(taskToRethrottle).get();
            list.rethrowFailures("subtasks");
            assertThat(finishedSubTasks + list.getTasks().size(), greaterThanOrEqualTo((long) request.request().getSlices()));
            assertThat(list.getTasks().size(), greaterThan(0));
        });
    }
    // Now rethrottle it so it'll finish
    // No throttle or "very fast"
    float newRequestsPerSecond = randomBoolean() ? Float.POSITIVE_INFINITY : between(1, 1000) * 100000;
    ListTasksResponse rethrottleResponse = rethrottle().setTaskId(taskToRethrottle).setRequestsPerSecond(newRequestsPerSecond).get();
    rethrottleResponse.rethrowFailures("Rethrottle");
    assertThat(rethrottleResponse.getTasks(), hasSize(1));
    BulkByScrollTask.Status status = (BulkByScrollTask.Status) rethrottleResponse.getTasks().get(0).getStatus();
    // Now check the resulting requests per second.
    if (request.request().getSlices() == 1) {
        // If there is a single slice it should match perfectly
        assertEquals(newRequestsPerSecond, status.getRequestsPerSecond(), Float.MIN_NORMAL);
    } else {
        /* Check that at least one slice was rethrottled. We won't always rethrottle all of them because they might have completed.
             * With multiple slices these numbers might not add up perfectly, thus the 1.01F. */
        long unfinished = status.getSliceStatuses().stream().filter(slice -> slice != null).filter(slice -> slice.getStatus().getTotal() > slice.getStatus().getSuccessfullyProcessed()).count();
        float maxExpectedSliceRequestsPerSecond = newRequestsPerSecond == Float.POSITIVE_INFINITY ? Float.POSITIVE_INFINITY : (newRequestsPerSecond / unfinished) * 1.01F;
        float minExpectedSliceRequestsPerSecond = newRequestsPerSecond == Float.POSITIVE_INFINITY ? Float.POSITIVE_INFINITY : (newRequestsPerSecond / request.request().getSlices()) * 0.99F;
        boolean oneSliceRethrottled = false;
        float totalRequestsPerSecond = 0;
        for (BulkByScrollTask.StatusOrException statusOrException : status.getSliceStatuses()) {
            if (statusOrException == null) {
                /* The slice can be null here because it was completed but hadn't reported its success back to the task when the
                     * rethrottle request came through. */
                continue;
            }
            assertNull(statusOrException.getException());
            BulkByScrollTask.Status slice = statusOrException.getStatus();
            if (slice.getTotal() > slice.getSuccessfullyProcessed()) {
                // This slice reports as not having completed so it should have been processed.
                assertThat(slice.getRequestsPerSecond(), both(greaterThanOrEqualTo(minExpectedSliceRequestsPerSecond)).and(lessThanOrEqualTo(maxExpectedSliceRequestsPerSecond)));
            }
            if (minExpectedSliceRequestsPerSecond <= slice.getRequestsPerSecond() && slice.getRequestsPerSecond() <= maxExpectedSliceRequestsPerSecond) {
                oneSliceRethrottled = true;
            }
            totalRequestsPerSecond += slice.getRequestsPerSecond();
        }
        assertTrue("At least one slice must be rethrottled", oneSliceRethrottled);
        /* Now assert that the parent request has the total requests per second. This is a much weaker assertion than that the parent
             * actually has the newRequestsPerSecond. For the most part it will. Sometimes it'll be greater because only unfinished requests
             * are rethrottled, the finished ones just keep whatever requests per second they had while they were running. But it might
             * also be less than newRequestsPerSecond because the newRequestsPerSecond is divided among running sub-requests and then the
             * requests are rethrottled. If one request finishes in between the division and the application of the new throttle then it
             * won't be rethrottled, thus only contributing its lower total. */
        assertEquals(totalRequestsPerSecond, status.getRequestsPerSecond(), totalRequestsPerSecond * 0.0001f);
    }
    // Now the response should come back quickly because we've rethrottled the request
    BulkByScrollResponse response = responseListener.get();
    assertThat("Entire request completed in a single batch. This may invalidate the test as throttling is done between batches.", response.getBatches(), greaterThanOrEqualTo(request.request().getSlices()));
}