1 Commits (b5ac6a0258be51a6d6c3a7743e498dc40014b64b)

Author	SHA1	Message	Date
Christian Haudum	27411cff08	Introduce worker queue in bloom gateway (#10976) ... Instead of calling the bloom store directly on each and every request to filter chunk refs based on the given filters, we want to queue requests in per-tenant queues and process batches of requests that can be multiplexed to avoid excessive seeking in the bloom block queriers when checking chunk matches. This PR re-uses the request queue implementation used in the query scheduler. To do so, it moves the queue related code from the scheduler into a separate package `pkg/queue` and renames occurrences of "querier" to "consumer" to be more generic. The bloom gateway instantiates the request queue when starting the service. The gRPC method `FilterChunkRefs` then enqueues incoming requests to that queue. **Special notes for your reviewer**: For testing purposes, this PR also contains a dummy implementation of the workers. The worker implementation - which includes multiplexing of multiple tasks - is subject to a separate PR. --------- Signed-off-by: Christian Haudum <christian.haudum@gmail.com>	2 years ago
Christian Haudum	9844fad8b4	Rename LeafQueue to TreeQueue (#8856) ... **What this PR does / why we need it**: `TreeQueue` is the semantically more correct term for this type of queue. Signed-off-by: Christian Haudum <christian.haudum@gmail.com>	2 years ago
Christian Haudum	be8b4eece3	Scheduler: Add query fairness control across multiple actors within a tenant (#8752) ... **What this PR does / why we need it**: This PR wires up the scheduler with the hierarchical queues. It is the last PR to implement https://github.com/grafana/loki/pull/8585. When these changes are in place, the client performing query requests can control their QoS (query fairness) using the `X-Actor-Path` HTTP header. This header controls in which sub-queue of the tenant's scheduler queue the query request is enqueued. The place within the hierarchy where it is enqueued defines the probability with which the request gets dequeued. A common use-case for this QoS control is giving each Grafana user within a tenant their fair share of query execution time. Any documentation is still missing and will be provided by follow-up PRs. **Special notes for your reviewer**: ```console $ gotest -count=1 -v ./pkg/scheduler/queue/... -test.run=TestQueryFairness === RUN TestQueryFairness === RUN TestQueryFairness/use_hierarchical_queues_=_false dequeue_qos_test.go:109: duration actor a 2.007765568s dequeue_qos_test.go:109: duration actor b 2.209088331s dequeue_qos_test.go:112: total duration 2.209280772s === RUN TestQueryFairness/use_hierarchical_queues_=_true dequeue_qos_test.go:109: duration actor b 605.283144ms dequeue_qos_test.go:109: duration actor a 2.270931324s dequeue_qos_test.go:112: total duration 2.271108551s --- PASS: TestQueryFairness (4.48s) --- PASS: TestQueryFairness/use_hierarchical_queues_=_false (2.21s) --- PASS: TestQueryFairness/use_hierarchical_queues_=_true (2.27s) PASS ok github.com/grafana/loki/pkg/scheduler/queue 4.491s ``` ```console $ gotest -count=5 -v ./pkg/scheduler/queue/... -bench=Benchmark -test.run=^$ -benchtime=10000x -benchmem goos: linux goarch: amd64 pkg: github.com/grafana/loki/pkg/scheduler/queue cpu: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz BenchmarkGetNextRequest BenchmarkGetNextRequest/without_sub-queues BenchmarkGetNextRequest/without_sub-queues-8 10000 29337 ns/op 1600 B/op 100 allocs/op BenchmarkGetNextRequest/without_sub-queues-8 10000 21348 ns/op 1600 B/op 100 allocs/op BenchmarkGetNextRequest/without_sub-queues-8 10000 21595 ns/op 1600 B/op 100 allocs/op BenchmarkGetNextRequest/without_sub-queues-8 10000 21189 ns/op 1600 B/op 100 allocs/op BenchmarkGetNextRequest/without_sub-queues-8 10000 21602 ns/op 1600 B/op 100 allocs/op BenchmarkGetNextRequest/with_1_level_of_sub-queues BenchmarkGetNextRequest/with_1_level_of_sub-queues-8 10000 33770 ns/op 2400 B/op 200 allocs/op BenchmarkGetNextRequest/with_1_level_of_sub-queues-8 10000 33596 ns/op 2400 B/op 200 allocs/op BenchmarkGetNextRequest/with_1_level_of_sub-queues-8 10000 34432 ns/op 2400 B/op 200 allocs/op BenchmarkGetNextRequest/with_1_level_of_sub-queues-8 10000 33760 ns/op 2400 B/op 200 allocs/op BenchmarkGetNextRequest/with_1_level_of_sub-queues-8 10000 33664 ns/op 2400 B/op 200 allocs/op BenchmarkGetNextRequest/with_2_levels_of_sub-queues BenchmarkGetNextRequest/with_2_levels_of_sub-queues-8 10000 71405 ns/op 3200 B/op 300 allocs/op BenchmarkGetNextRequest/with_2_levels_of_sub-queues-8 10000 59472 ns/op 3200 B/op 300 allocs/op BenchmarkGetNextRequest/with_2_levels_of_sub-queues-8 10000 117163 ns/op 3200 B/op 300 allocs/op BenchmarkGetNextRequest/with_2_levels_of_sub-queues-8 10000 106505 ns/op 3200 B/op 300 allocs/op BenchmarkGetNextRequest/with_2_levels_of_sub-queues-8 10000 64374 ns/op 3200 B/op 300 allocs/op BenchmarkQueueRequest BenchmarkQueueRequest-8 10000 168391 ns/op 320588 B/op 1156 allocs/op BenchmarkQueueRequest-8 10000 166203 ns/op 320587 B/op 1156 allocs/op BenchmarkQueueRequest-8 10000 149518 ns/op 320584 B/op 1156 allocs/op BenchmarkQueueRequest-8 10000 219776 ns/op 320583 B/op 1156 allocs/op BenchmarkQueueRequest-8 10000 185198 ns/op 320597 B/op 1156 allocs/op PASS ok github.com/grafana/loki/pkg/scheduler/queue 64.648s ``` Signed-off-by: Christian Haudum <christian.haudum@gmail.com>	2 years ago
Christian Haudum	22b672986a	Revert "Implement hierarchical queues for query scheduler" (#8796) ... Reverts grafana/loki#8691 These changes caused problems with the scheduler in our dev environments. I will conduct further testing and include the reverted changes in https://github.com/grafana/loki/pull/8752	2 years ago
Christian Haudum	4b74a5b815	Implement hierarchical queues for query scheduler (#8691) ... ### What this PR does / why we need it: This PR is the first step towards hierarchical queues within the query scheduler. Hierarchical queues can be used to ensure query fairness across multiple actors within a single tenant, such as individual (human) users using the same tenant. The change is described in the LID ["Query fairness across users within tenants"](https://grafana.com/docs/loki/next/lids/0003-queryfairnessinscheduler/). * Extract metrics for `RequestQueue` into their own `metrics` field, so that they can extended and passed more easily in the future. * Move mapping for tenant queues into its own data structure (ordered map) so its functionality can be reused. It also simplifies the code inside the `tenantQueues` struct. * Add the `LeafQueue` struct that implements hierarchical queues. Sub-queues are dequeued in a round-robin manner using the same implementation of `Mapping` that is also used by the tenant queues. ### Special notes for your reviewer Part of https://github.com/grafana/loki/pull/8585 Contains the changes from PR https://github.com/grafana/loki/pull/8722 **This PR does not yet make use of the new hierarchical queues, it only adds the relevant data structures. Wiring up will be done in a follow up PR.** While the new `Mapping` data structure adds an unsignificant overhead to the time/op I consider the simplification an overall better tradeoff. ```console $ benchstat scheduler-bench-old.txt scheduler-bench-new.txt name old time/op new time/op delta GetNextRequest-8 18.0µs ± 2% 20.5µs ± 3% +13.89% (p=0.008 n=5+5) QueueRequest-8 56.6µs ± 0% 53.5µs ± 3% -5.48% (p=0.008 n=5+5) name old alloc/op new alloc/op delta GetNextRequest-8 1.60kB ± 0% 1.60kB ± 0% ~ (all equal) QueueRequest-8 35.6kB ± 0% 29.8kB ± 0% -16.37% (p=0.008 n=5+5) name old allocs/op new allocs/op delta GetNextRequest-8 100 ± 0% 100 ± 0% ~ (all equal) QueueRequest-8 819 ± 0% 806 ± 0% -1.59% (p=0.008 n=5+5) ``` #### Complexity of operations | Operation | Function | Complexity | | --------- | -------- | ---------- | | Insert | `Put(key string, value *tenantQueue) bool` | `O(1)`| | Select | `GetNext(idx QueueIndex) *tenantQueue` | `O(1)` (best case) `O(n)` (worst case) | | Remove | `Remove(key string) bool` | `O(1)` | --- Signed-off-by: Christian Haudum <christian.haudum@gmail.com>	2 years ago