Mimir query engine: add memory consumption per query limit

CHANGELOG.md

@@ -15,7 +15,7 @@
 * [FEATURE] Continuous-test: now runable as a module with `mimir -target=continuous-test`. #7747
 * [FEATURE] Store-gateway: Allow specific tenants to be enabled or disabled via `-store-gateway.enabled-tenants` or `-store-gateway.disabled-tenants` CLI flags or their corresponding YAML settings. #7653
 * [FEATURE] New `-<prefix>.s3.bucket-lookup-type` flag configures lookup style type, used to access bucket in s3 compatible providers. #7684
-* [FEATURE] Querier: add experimental streaming PromQL engine, enabled with `-querier.promql-engine=streaming`. #7693 #7898 #7899 #8023 #8058 #8096 #8121 #8197
+* [FEATURE] Querier: add experimental streaming PromQL engine, enabled with `-querier.promql-engine=streaming`. #7693 #7898 #7899 #8023 #8058 #8096 #8121 #8197 #8230
 * [FEATURE] New `/ingester/unregister-on-shutdown` HTTP endpoint allows dynamic access to ingesters' `-ingester.ring.unregister-on-shutdown` configuration. #7739
 * [FEATURE] Server: added experimental [PROXY protocol support](https://www.haproxy.org/download/2.3/doc/proxy-protocol.txt). The PROXY protocol support can be enabled via `-server.proxy-protocol-enabled=true`. When enabled, the support is added both to HTTP and gRPC listening ports. #7698
 * [FEATURE] mimirtool: Add `runtime-config verify` sub-command, for verifying Mimir runtime config files. #8123

cmd/mimir/config-descriptor.json

@@ -3545,6 +3545,17 @@
           "fieldFlag": "querier.max-fetched-chunk-bytes-per-query",
           "fieldType": "int"
         },
+        {
+          "kind": "field",
+          "name": "max_estimated_memory_consumption_per_query",
+          "required": false,
+          "desc": "The maximum estimated memory a single query can consume at once, in bytes. This limit is only enforced when Mimir's query engine is in use. This limit is enforced in the querier. 0 to disable.",
+          "fieldValue": null,
+          "fieldDefaultValue": 0,
+          "fieldFlag": "querier.max-estimated-memory-consumption-per-query",
+          "fieldType": "int",
+          "fieldCategory": "experimental"
+        },
         {
           "kind": "field",
           "name": "max_query_lookback",

cmd/mimir/help-all.txt.tmpl

@@ -1737,6 +1737,8 @@ Usage of ./cmd/mimir/mimir:
     	The number of workers running in each querier process. This setting limits the maximum number of concurrent queries in each querier. (default 20)
   -querier.max-estimated-fetched-chunks-per-query-multiplier float
     	[experimental] Maximum number of chunks estimated to be fetched in a single query from ingesters and store-gateways, as a multiple of -querier.max-fetched-chunks-per-query. This limit is enforced in the querier. Must be greater than or equal to 1, or 0 to disable.
+  -querier.max-estimated-memory-consumption-per-query uint
+    	[experimental] The maximum estimated memory a single query can consume at once, in bytes. This limit is only enforced when Mimir's query engine is in use. This limit is enforced in the querier. 0 to disable.
   -querier.max-fetched-chunk-bytes-per-query int
     	The maximum size of all chunks in bytes that a query can fetch from ingesters and store-gateways. This limit is enforced in the querier and ruler. 0 to disable.
   -querier.max-fetched-chunks-per-query int

docs/sources/mimir/configure/about-versioning.md

@@ -131,6 +131,7 @@ The following features are currently experimental:
   - Enable PromQL experimental functions (`-querier.promql-experimental-functions-enabled`)
   - Allow streaming of `/active_series` responses to the frontend (`-querier.response-streaming-enabled`)
   - Streaming PromQL engine (`-querier.promql-engine=streaming` and `-querier.enable-promql-engine-fallback`)
+  - Maximum estimated memory consumption per query limit (`-querier.max-estimated-memory-consumption-per-query`)
 - Query-frontend
   - `-query-frontend.querier-forget-delay`
   - Instant query splitting (`-query-frontend.split-instant-queries-by-interval`)

docs/sources/mimir/configure/configuration-parameters/index.md

@@ -3164,6 +3164,12 @@ The `limits` block configures default and per-tenant limits imposed by component
 # CLI flag: -querier.max-fetched-chunk-bytes-per-query
 [max_fetched_chunk_bytes_per_query: <int> | default = 0]
 
+# (experimental) The maximum estimated memory a single query can consume at
+# once, in bytes. This limit is only enforced when Mimir's query engine is in
+# use. This limit is enforced in the querier. 0 to disable.
+# CLI flag: -querier.max-estimated-memory-consumption-per-query
+[max_estimated_memory_consumption_per_query: <int> | default = 0]
+
 # Limit how long back data (series and metadata) can be queried, up until
 # <lookback> duration ago. This limit is enforced in the query-frontend, querier
 # and ruler. If the requested time range is outside the allowed range, the

docs/sources/mimir/manage/mimir-runbooks/_index.md

@@ -1992,6 +1992,23 @@ How to **fix** it:
 - Consider reducing the time range and/or cardinality of the query. To reduce the cardinality of the query, you can add more label matchers to the query, restricting the set of matching series.
 - Consider increasing the per-tenant limit by using the `-querier.max-fetched-chunk-bytes-per-query` option (or `max_fetched_chunk_bytes_per_query` in the runtime configuration).
 
+### err-mimir-max-estimated-memory-consumption-per-query
+
+This error occurs when execution of a query exceeds the limit on the maximum estimated memory consumed by a single query.
+
+This limit is used to protect the system’s stability from potential abuse or mistakes, when running a query fetching a huge amount of data.
+This limit only applies when Mimir's query engine is used (ie. `-querier.promql-engine=streaming`).
+To configure the limit on a global basis, use the `-querier.max-estimated-memory-consumption-per-query` option.
+To configure the limit on a per-tenant basis, set the `max_estimated_memory_consumption_per_query` per-tenant override in the runtime configuration.
+
+How to **fix** it:
+
+- Consider reducing the time range of the query.
+- Consider reducing the cardinality of the query. To reduce the cardinality of the query, you can add more label matchers to the query, restricting the set of matching series.
+- Consider applying aggregations such as `sum` or `avg` to the query.
+- Consider increasing the global limit by using the `-querier.max-estimated-memory-consumption-per-query` option.
+- Consider increasing the limit on a per-tenant basis by using the `max_estimated_memory_consumption_per_query` per tenant-override in the runtime configuration.
+
 ### err-mimir-max-query-length
 
 This error occurs when the time range of a partial (after possible splitting, sharding by the query-frontend) query exceeds the configured maximum length. For a limit on the total query length, see [err-mimir-max-total-query-length](#err-mimir-max-total-query-length).

pkg/querier/querier.go

@@ -164,7 +164,8 @@ func New(cfg Config, limits *validation.Overrides, distributor Distributor, stor
 	case standardPromQLEngine:
 		eng = promql.NewEngine(opts)
 	case streamingPromQLEngine:
-		streamingEngine, err := streamingpromql.NewEngine(opts)
+		limitsProvider := &tenantQueryLimitsProvider{limits: limits}
+		streamingEngine, err := streamingpromql.NewEngine(opts, limitsProvider)
 		if err != nil {
 			return nil, nil, nil, err
 		}
@@ -623,3 +624,16 @@ func logClampEvent(spanLog *spanlogger.SpanLogger, originalT, clampedT int64, mi
 		"updated", util.TimeFromMillis(clampedT).String(),
 	)
 }
+
+type tenantQueryLimitsProvider struct {
+	limits *validation.Overrides
+}
+
+func (p *tenantQueryLimitsProvider) GetMaxEstimatedMemoryConsumptionPerQuery(ctx context.Context) (uint64, error) {
+	tenantID, err := tenant.TenantID(ctx)
+	if err != nil {
+		return 0, err
+	}
+
+	return p.limits.MaxEstimatedMemoryConsumptionPerQuery(tenantID), nil
+}

pkg/streamingpromql/benchmarks/comparison_test.go

@@ -44,7 +44,7 @@ func BenchmarkQuery(b *testing.B) {
 
 	opts := streamingpromql.NewTestEngineOpts()
 	prometheusEngine := promql.NewEngine(opts)
-	streamingEngine, err := streamingpromql.NewEngine(opts)
+	streamingEngine, err := streamingpromql.NewEngine(opts, streamingpromql.NewStaticQueryLimitsProvider(0))
 	require.NoError(b, err)
 
 	// Important: the names below must remain in sync with the names used in tools/benchmark-query-engine.
@@ -96,7 +96,7 @@ func TestBothEnginesReturnSameResultsForBenchmarkQueries(t *testing.T) {
 
 	opts := streamingpromql.NewTestEngineOpts()
 	prometheusEngine := promql.NewEngine(opts)
-	streamingEngine, err := streamingpromql.NewEngine(opts)
+	streamingEngine, err := streamingpromql.NewEngine(opts, streamingpromql.NewStaticQueryLimitsProvider(0))
 	require.NoError(t, err)
 
 	ctx := user.InjectOrgID(context.Background(), UserID)
@@ -123,7 +123,7 @@ func TestBenchmarkSetup(t *testing.T) {
 	q := createBenchmarkQueryable(t, []int{1})
 
 	opts := streamingpromql.NewTestEngineOpts()
-	streamingEngine, err := streamingpromql.NewEngine(opts)
+	streamingEngine, err := streamingpromql.NewEngine(opts, streamingpromql.NewStaticQueryLimitsProvider(0))
 	require.NoError(t, err)
 
 	ctx := user.InjectOrgID(context.Background(), UserID)

pkg/streamingpromql/engine.go

@@ -17,7 +17,7 @@ import (
 
 const defaultLookbackDelta = 5 * time.Minute // This should be the same value as github.com/prometheus/prometheus/promql.defaultLookbackDelta.
 
-func NewEngine(opts promql.EngineOpts) (promql.QueryEngine, error) {
+func NewEngine(opts promql.EngineOpts, limitsProvider QueryLimitsProvider) (promql.QueryEngine, error) {
 	lookbackDelta := opts.LookbackDelta
 	if lookbackDelta == 0 {
 		lookbackDelta = defaultLookbackDelta
@@ -36,21 +36,23 @@ func NewEngine(opts promql.EngineOpts) (promql.QueryEngine, error) {
 	}
 
 	return &Engine{
-		lookbackDelta: lookbackDelta,
-		timeout:       opts.Timeout,
+		lookbackDelta:  lookbackDelta,
+		timeout:        opts.Timeout,
+		limitsProvider: limitsProvider,
 	}, nil
 }
 
 type Engine struct {
-	lookbackDelta time.Duration
-	timeout       time.Duration
+	lookbackDelta  time.Duration
+	timeout        time.Duration
+	limitsProvider QueryLimitsProvider
 }
 
-func (e *Engine) NewInstantQuery(_ context.Context, q storage.Queryable, opts promql.QueryOpts, qs string, ts time.Time) (promql.Query, error) {
-	return newQuery(q, opts, qs, ts, ts, 0, e)
+func (e *Engine) NewInstantQuery(ctx context.Context, q storage.Queryable, opts promql.QueryOpts, qs string, ts time.Time) (promql.Query, error) {
+	return newQuery(ctx, q, opts, qs, ts, ts, 0, e)
 }
 
-func (e *Engine) NewRangeQuery(_ context.Context, q storage.Queryable, opts promql.QueryOpts, qs string, start, end time.Time, interval time.Duration) (promql.Query, error) {
+func (e *Engine) NewRangeQuery(ctx context.Context, q storage.Queryable, opts promql.QueryOpts, qs string, start, end time.Time, interval time.Duration) (promql.Query, error) {
 	if interval <= 0 {
 		return nil, fmt.Errorf("%v is not a valid interval for a range query, must be greater than 0", interval)
 	}
@@ -59,5 +61,27 @@ func (e *Engine) NewRangeQuery(_ context.Context, q storage.Queryable, opts prom
 		return nil, fmt.Errorf("range query time range is invalid: end time %v is before start time %v", end.Format(time.RFC3339), start.Format(time.RFC3339))
 	}
 
-	return newQuery(q, opts, qs, start, end, interval, e)
+	return newQuery(ctx, q, opts, qs, start, end, interval, e)
+}
+
+type QueryLimitsProvider interface {
+	// GetMaxEstimatedMemoryConsumptionPerQuery returns the maximum estimated memory allowed to be consumed by a query in bytes, or 0 to disable the limit.
+	GetMaxEstimatedMemoryConsumptionPerQuery(ctx context.Context) (uint64, error)
+}
+
+// NewStaticQueryLimitsProvider returns a QueryLimitsProvider that always returns the provided limits.
+//
+// This should generally only be used in tests.
+func NewStaticQueryLimitsProvider(maxEstimatedMemoryConsumptionPerQuery uint64) QueryLimitsProvider {
+	return staticQueryLimitsProvider{
+		maxEstimatedMemoryConsumptionPerQuery: maxEstimatedMemoryConsumptionPerQuery,
+	}
+}
+
+type staticQueryLimitsProvider struct {
+	maxEstimatedMemoryConsumptionPerQuery uint64
+}
+
+func (p staticQueryLimitsProvider) GetMaxEstimatedMemoryConsumptionPerQuery(_ context.Context) (uint64, error) {
+	return p.maxEstimatedMemoryConsumptionPerQuery, nil
 }