Table of Contents

Querier/Query #

The thanos query command (also known as “Querier”) implements the Prometheus HTTP v1 API to query data in a Thanos cluster via PromQL.

In short, it gathers the data needed to evaluate the query from underlying StoreAPIs, evaluates the query and returns the result.

Querier is fully stateless and horizontally scalable.

Example command to run Querier:

thanos query \
    --http-address     "" \
    --store            "<store-api>:<grpc-port>" \
    --store            "<store-api2>:<grpc-port>"

Querier use cases, why do I need this component? #

Thanos Querier essentially allows to aggregate and optionally deduplicate multiple metrics backends under single Prometheus Query endpoint.

Global View #

Since for Querier “a backend” is anything that implements gRPC StoreAPI we can aggregate data from any number of the different storages like:

  • Prometheus (see Sidecar)
  • Object Storage (see Store Gateway)
  • Global alerting/recording rules evaluations (see Ruler)
  • Metrics received from Prometheus remote write streams (see Receiver)
  • Another Querier (you can stack Queriers on top of each other)
  • Non-Prometheus systems!

Thanks to that, you can run queries (manually, from Grafana or via Alerting rule) that aggregate metrics from mix of those sources.

Some examples:

  • sum(cpu_used{cluster=~"cluster-(eu1|eu2|eu3|us1|us2|us3)", job="service1"}) that will give you sum of CPU used inside all listed clusters for service service1. This will work even if those clusters runs multiple Prometheus servers each. Querier will know which data sources to query.

  • In single cluster you shard Prometheus functionally or have different Prometheus instances for different tenants. You can spin up Querier to have access to both within single Query evaluation.

Run-time deduplication of HA groups #

Prometheus is stateful and does not allow replicating its database. This means that increasing high availability by running multiple Prometheus replicas is not very easy to use. Simple load balancing will not work as for example after some crash, replica might be up but querying such replica will result in small gap during the period it was down. You have a second replica that maybe was up, but it could be down in other moment (e.g rolling restart), so load balancing on top of those is not working well.

Thanos Querier instead pulls the data from both replicas, and deduplicate those signals, filling the gaps if any, transparently to the Querier consumer.

Metric Query Flow Overview #


Overall QueryAPI exposed by Thanos is guaranteed to be compatible with Prometheus 2.x. API. The above diagram shows what Querier does for each Prometheus query request.

See here on how to connect Querier with desired StoreAPIs.

Deduplication #

The query layer can deduplicate series that were collected from high-availability pairs of data sources such as Prometheus. A fixed single or multiple replica labels must be chosen for the entire cluster and can then be passed to query nodes on startup.

Two or more series that are only distinguished by the given replica label, will be merged into a single time series. This also hides gaps in collection of a single data source.

An example with a single replica labels: #

  • Prometheus + sidecar “A”: cluster=1,env=2,replica=A
  • Prometheus + sidecar “B”: cluster=1,env=2,replica=B
  • Prometheus + sidecar “A” in different cluster: cluster=2,env=2,replica=A

If we configure Querier like this:

thanos query \
    --http-address        "" \
    --query.replica-label "replica" \
    --store               "<store-api>:<grpc-port>" \
    --store               "<store-api2>:<grpc-port>" \

And we query for metric up{job="prometheus",env="2"} with this option we will get 2 results:

  • up{job="prometheus",env="2",cluster="1"} 1
  • up{job="prometheus",env="2",cluster="2"} 1

WITHOUT this replica flag (deduplication turned off), we will get 3 results:

  • up{job="prometheus",env="2",cluster="1",replica="A"} 1
  • up{job="prometheus",env="2",cluster="1",replica="B"} 1
  • up{job="prometheus",env="2",cluster="2",replica="A"} 1

The same output will be present for this example with multiple replica labels: #

  • Prometheus + sidecar “A”: cluster=1,env=2,replica=A,replicaX=A
  • Prometheus + sidecar “B”: cluster=1,env=2,replica=B,replicaX=B
  • Prometheus + sidecar “A” in different cluster: cluster=2,env=2,replica=A,replicaX=A
thanos query \
    --http-address        "" \
    --query.replica-label "replica" \
    --query.replica-label "replicaX" \
    --store               "<store-api>:<grpc-port>" \
    --store               "<store-api2>:<grpc-port>" \

This logic can also be controlled via parameter on QueryAPI. More details below.

Query API Overview #

As mentioned, Query API exposed by Thanos is guaranteed to be compatible with Prometheus 2.x. API. However for additional Thanos features on top of Prometheus, Thanos adds:

  • partial response behaviour
  • several additional parameters listed below
  • custom response fields.

Let’s walk through all of those extensions:

Partial Response #

QueryAPI and StoreAPI has additional behaviour controlled via query parameter called PartialResponseStrategy.

This parameter controls tradeoff between accuracy and availability.

Partial response is a potentially missed result within query against QueryAPI or StoreAPI. This can happen if one of StoreAPIs is returning error or timeout whereas couple of others returns success. It does not mean you are missing data, you might lucky enough that you actually get the correct data as the broken StoreAPI did not have anything for your query.

If partial response happen QueryAPI returns human readable warnings explained here.

Now it supports two strategies:

  • “warn”
  • “abort” (default)

NOTE: Having a warning does not necessarily mean partial response (e.g no store matched query warning).

Querier also allows to configure different timeouts:

  • --query.timeout
  • --store.response-timeout

If you prefer availability over accuracy you can set tighter timeout to underlying StoreAPI than overall query timeout. If partial response strategy is NOT abort, this will “ignore” slower StoreAPIs producing just warning with 200 status code response.

Deduplication replica labels. #

HTTP URL/FORM parameterTypeDefaultExample
replicaLabels[]stringquery.replica-label flag (default: empty).replicaLabels=replicaA&replicaLabels=replicaB

This overwrites the query.replica-label cli flag to allow dynamic replica labels at query time.

Deduplication Enabled #

HTTP URL/FORM parameterTypeDefaultExample
dedupBooleanTrue, but effect depends on query.replica configuration flag.1, t, T, TRUE, true, True for “True”

This controls if query results should be deduplicated using the replica labels.

Auto downsampling #

HTTP URL/FORM parameterTypeDefaultExample
max_source_resolutionFloat64/time.Duration/model.Durationstep / 5 or 0 if is false (default: False)5m

Max source resolution is max resolution in seconds we want to use for data we query for.

Available options:

  • auto - Select downsample resolution automatically based on the query.
  • 0 - Only use raw data.
  • 5m - Use max 5m downsampling.
  • 1h - Use max 1h downsampling.

Partial Response Strategy #

// TODO(bwplotka): Update. This will change to “strategy” soon as PartialResponseStrategy enum here

HTTP URL/FORM parameterTypeDefaultExample
partial_responseBooleanquery.partial-response flag (default: True)1, t, T, TRUE, true, True for “True”

If true, then all storeAPIs that will be unavailable (and thus return no data) will not cause query to fail, but instead return warning.

Custom Response Fields #

Any additional field does not break compatibility, however there is no guarantee that Grafana or any other client will understand those.

Currently Thanos UI exposed by Thanos understands

type queryData struct {
	ResultType promql.ValueType `json:"resultType"`
	Result     promql.Value     `json:"result"`

	// Additional Thanos Response field.
	Warnings []error `json:"warnings,omitempty"`

Additional field is Warnings that contains every error that occurred that is assumed non critical. partial_response option controls if storeAPI unavailability is considered critical.

Concurrent Selects #

Thanos Querier has the ability to perform concurrent select request per query. It dissects given PromQL statement and executes selectors concurrently against the discovered StoreAPIs. The maximum number of concurrent requests are being made per query is controlled by query.max-concurrent-select flag. Keep in mind that the maximum number of concurrent queries that are handled by querier is controlled by query.max-concurrent. Please consider implications of combined value while tuning the querier.

Store filtering #

It’s possible to provide a set of matchers to the Querier api to select specific stores to be used during the query using the storeMatch[] parameter. It is useful when debugging a slow/broken store. It uses the same format as the matcher of Prometheus’ federate api. Note that at the moment the querier only supports the __address__ which contain the address of the store as it is shown on the /stores endpoint of the UI.


- targets:
  - prometheus-foo.thanos-sidecar:10901
  - prometheus-bar.thanos-sidecar:10901

Will only return metrics from prometheus-foo.thanos-sidecar:10901

Expose UI on a sub-path #

It is possible to expose thanos-query UI and optionally API on a sub-path. The sub-path can be defined either statically or dynamically via an HTTP header. Static path prefix definition follows the pattern used in Prometheus, where web.route-prefix option defines HTTP request path prefix (endpoints prefix) and web.external-prefix prefixes the URLs in HTML code and the HTTP redirect responses.

Additionally, Thanos supports dynamic prefix configuration, which is not yet implemented by Prometheus. Dynamic prefixing simplifies setup when thanos query is exposed on a sub-path behind a reverse proxy, for example, via a Kubernetes ingress controller Traefik or nginx. If PathPrefixStrip: /some-path option or traefik.frontend.rule.type: PathPrefixStrip Kubernetes Ingress annotation is set, then Traefik writes the stripped prefix into X-Forwarded-Prefix header. Then, thanos query --web.prefix-header=X-Forwarded-Prefix will serve correct HTTP redirects and links prefixed by the stripped path.

File SD # flag provides a path to a JSON or YAML formatted file, which contains a list of targets in Prometheus target format.

Example file SD file in YAML:

- targets:
  - prometheus-0.thanos-sidecar:10901
  - prometheus-1.thanos-sidecar:10901
  - thanos-store:10901
  - thanos-short-store:10901
  - thanos-rule:10901
- targets:
  - prometheus-0.thanos-sidecar.infra:10901
  - prometheus-1.thanos-sidecar.infra:10901
  - thanos-store.infra:10901

Active Query Tracking # is an option which allows the user to specify a directory which will contain a file to track active queries. To enable this feature, the user has to specify a directory other than “”, since that is skipped being the default.

Flags #

usage: thanos query [<flags>]

Query node exposing PromQL enabled Query API with data retrieved from multiple
store nodes.

                                 The external Thanos Query URL that would be set
                                 in all alerts 'Source' field.
      --enable-feature= ...      Comma separated experimental feature names
                                 to enable.The current list of features is
      --endpoint=<endpoint> ...  Addresses of statically configured Thanos
                                 API servers (repeatable). The scheme may be
                                 prefixed with 'dns+' or 'dnssrv+' to detect
                                 Thanos API servers through respective DNS
      --endpoint-strict=<staticendpoint> ...
                                 Addresses of only statically configured Thanos
                                 API servers that are always used, even if
                                 the health check fails. Useful if you have a
                                 caching layer on top.
                                 Listen ip:port address for gRPC endpoints
                                 (StoreAPI). Make sure this address is routable
                                 from other components.
                                 Server name to verify the hostname on
                                 the returned gRPC certificates. See
      --grpc-client-tls-ca=""    TLS CA Certificates to use to verify gRPC
      --grpc-client-tls-cert=""  TLS Certificates to use to identify this client
                                 to the server
      --grpc-client-tls-key=""   TLS Key for the client's certificate
      --grpc-client-tls-secure   Use TLS when talking to the gRPC server
                                 Disable TLS certificate verification i.e self
                                 signed, signed by fake CA
      --grpc-compression=none    Compression algorithm to use for gRPC requests
                                 to other clients. Must be one of: snappy, none
      --grpc-grace-period=2m     Time to wait after an interrupt received for
                                 GRPC Server.
                                 The grpc server max connection age.
                                 This controls how often to re-read the tls
                                 certificates and redo the TLS handshake
      --grpc-server-tls-cert=""  TLS Certificate for gRPC server, leave blank to
                                 disable TLS
                                 TLS CA to verify clients against. If no
                                 client CA is specified, there is no client
                                 verification on server side. (tls.NoClientCert)
      --grpc-server-tls-key=""   TLS Key for the gRPC server, leave blank to
                                 disable TLS
  -h, --help                     Show context-sensitive help (also try
                                 --help-long and --help-man).
                                 Listen host:port for HTTP endpoints.
      --http-grace-period=2m     Time to wait after an interrupt received for
                                 HTTP Server.
      --http.config=""           [EXPERIMENTAL] Path to the configuration file
                                 that can enable TLS or authentication for all
                                 HTTP endpoints.
      --log.format=logfmt        Log format to use. Possible options: logfmt or
      --log.level=info           Log filtering level.
      --log.request.decision=    Deprecation Warning - This flag would
                                 be soon deprecated, and replaced with
                                 `request.logging-config`. Request Logging
                                 for logging the start and end of requests. By
                                 default this flag is disabled. LogFinishCall:
                                 Logs the finish call of the requests.
                                 LogStartAndFinishCall: Logs the start and
                                 finish call of the requests. NoLogCall: Disable
                                 request logging.""
                                 Directory to log currently active queries in
                                 the file.  Enable automatic adjustment (step / 5) to what
                                 source of data should be used in store gateways
                                 if no max_source_resolution param is specified.
                                 Set default evaluation interval for sub
      --query.default-step=1s    Set default step for range queries. Default
                                 step is only used when step is not set in UI.
                                 In such cases, Thanos UI will use default
                                 step to calculate resolution (resolution
                                 = max(rangeSeconds / 250, defaultStep)).
                                 This will not work from Grafana, but Grafana
                                 has __step variable which can be used.
                                 The maximum lookback duration for retrieving
                                 metrics during expression evaluations.
                                 PromQL always evaluates the query for the
                                 certain timestamp (query range timestamps are
                                 deduced by step). Since scrape intervals might
                                 be different, PromQL looks back for given
                                 amount of time to get latest sample. If it
                                 exceeds the maximum lookback delta it assumes
                                 series is stale and returns none (a gap).
                                 This is why lookback delta should be set to at
                                 least 2 times of the slowest scrape interval.
                                 If unset it will use the promql default of 5m.
      --query.max-concurrent=20  Maximum number of queries processed
                                 concurrently by query node.
                                 Maximum number of select requests made
                                 concurrently per a query.
                                 The default metadata time range duration for
                                 retrieving labels through Labels and Series API
                                 when the range parameters are not specified.
                                 The zero value means range covers the time
                                 since the beginning.
      --query.partial-response   Enable partial response for queries if
                                 no partial_response param is specified.
                                 --no-query.partial-response for disabling.
      --query.replica-label=QUERY.REPLICA-LABEL ...
                                 Labels to treat as a replica indicator along
                                 which data is deduplicated. Still you will
                                 be able to query without deduplication using
                                 'dedup=false' parameter. Data includes time
                                 series, recording rules, and alerting rules.
      --query.telemetry.request-duration-seconds-quantiles=0.1... ...
                                 The quantiles for exporting metrics about the
                                 request duration quantiles.
      --query.telemetry.request-samples-quantiles=100... ...
                                 The quantiles for exporting metrics about the
                                 samples count quantiles.
      --query.telemetry.request-series-seconds-quantiles=10... ...
                                 The quantiles for exporting metrics about the
                                 series count quantiles.
      --query.timeout=2m         Maximum time to process query by query node.
                                 Alternative to 'request.logging-config-file'
                                 flag (mutually exclusive). Content
                                 of YAML file with request logging
                                 configuration. See format details:
                                 Path to YAML file with request logging
                                 configuration. See format details:
      --selector-label=<name>="<value>" ...
                                 Query selector labels that will be exposed in
                                 info endpoint (repeated).
      --store=<store> ...        Deprecation Warning - This flag is deprecated
                                 and replaced with `endpoint`. Addresses of
                                 statically configured store API servers
                                 (repeatable). The scheme may be prefixed with
                                 'dns+' or 'dnssrv+' to detect store API servers
                                 through respective DNS lookups.
      --store-strict=<staticstore> ...
                                 Deprecation Warning - This flag is deprecated
                                 and replaced with `endpoint-strict`. Addresses
                                 of only statically configured store API servers
                                 that are always used, even if the health check
                                 fails. Useful if you have a caching layer on
                                 If a Store doesn't send any data in this
                                 specified duration then a Store will be ignored
                                 and partial data will be returned if it's
                                 enabled. 0 disables timeout.
                                 Interval between DNS resolutions.<path> ...
                                 Path to files that contain addresses of store
                                 API servers. The path can be a glob pattern
                                 (repeatable).     Refresh interval to re-read file SD files.
                                 It is used as a resync fallback.
                                 Timeout before an unhealthy store is cleaned
                                 from the store UI page.
                                 Alternative to 'tracing.config-file' flag
                                 (mutually exclusive). Content of YAML file
                                 with tracing configuration. See format details:
                                 Path to YAML file with tracing
                                 configuration. See format details:
      --version                  Show application version.
      --web.disable-cors         Whether to disable CORS headers to be set by
                                 Thanos. By default Thanos sets CORS headers to
                                 be allowed by all.
      --web.external-prefix=""   Static prefix for all HTML links and
                                 redirect URLs in the UI query web interface.
                                 Actual endpoints are still served on / or the
                                 web.route-prefix. This allows thanos UI to be
                                 served behind a reverse proxy that strips a URL
      --web.prefix-header=""     Name of HTTP request header used for dynamic
                                 prefixing of UI links and redirects.
                                 This option is ignored if web.external-prefix
                                 argument is set. Security risk: enable
                                 this option only if a reverse proxy in
                                 front of thanos is resetting the header.
                                 The --web.prefix-header=X-Forwarded-Prefix
                                 option can be useful, for example, if Thanos
                                 UI is served via Traefik reverse proxy with
                                 PathPrefixStrip option enabled, which sends the
                                 stripped prefix value in X-Forwarded-Prefix
                                 header. This allows thanos UI to be served on a
      --web.route-prefix=""      Prefix for API and UI endpoints. This allows
                                 thanos UI to be served on a sub-path.
                                 Defaults to the value of --web.external-prefix.
                                 This option is analogous to --web.route-prefix
                                 of Prometheus.

Exported metrics #

Thanos Query also exports metrics about its own performance. You can find a list with these metrics below.

Disclaimer: this list is incomplete. The remaining metrics will be added over time.

thanos_store_api_query_duration_secondsHistogramsamples_le, series_leDuration of the Thanos Store API select phase for a query according to the amount of samples and series selected.

Found a typo, inconsistency or missing information in our docs? Help us to improve Thanos documentation by proposing a fix on GitHub here ❤️