The prom-label-proxy can enforce a given label in a given PromQL query, in Prometheus API responses or in Alertmanager API requests. As an example (but not only), this allows read multi-tenancy for projects like Prometheus, Alertmanager or Thanos.
This proxy does not perform authentication or authorization, this has to happen before the request reaches this proxy, allowing you to use any authN/authZ system you want. The kube-rbac-proxy is an example for such an additional building block. Additionally, you can use prom-label-proxy as a library in your own proxy, like what is done in prom-authzed-proxy.
It's not a goal for this project to solve write tenant isolation for multi-tenant Prometheus:
- If a tenant controls its scrape target configuration the tenant can set arbitrary labels via its relabelling configuration, thereby being able to pollute other tenant's metrics.
- If the ingestion configuration honor_labels is set for a tenant's target, that target can pollute other tenant's metrics as Prometheus respects any labels exposed by the target.
See Prometheus Operator label enforcement, Thanos soft/hard tenancy or Cortex as example solution to that.
See: https://github.com/prometheus-community/helm-charts/tree/main/charts/prom-label-proxy
We publish docker images for each release, see:
quay.io/prometheuscommunity/prom-label-proxyfor newest imagesquay.io/coreos/prom-label-proxy:v0.1.0for the initial v0.1.0 release.
If you want to build prom-label-proxy from source you would need a working installation of
the Go 1.15+ toolchain (GOPATH, PATH=${GOPATH}/bin:${PATH}).
prom-label-proxy can be downloaded and built by running:
go get github.com/prometheus-community/prom-label-proxyThis application proxies the following endpoints and it ensures that a particular label is enforced in the particular request and response:
/federatefor GET method (Prometheus)/api/v1/query_exemplarsfor GET and POST methods (Prometheus/Thanos)/api/v1/queryfor GET and POST methods (Prometheus/Thanos)/api/v1/query_rangefor GET and POST methods (Prometheus/Thanos)/api/v1/seriesfor GET method (Prometheus/Thanos)/api/v1/rulesfor GET method (Prometheus/Thanos)/api/v1/alertsfor GET method (Prometheus/Thanos)/api/v2/silencesfor GET and POST methods (Alertmanager)/api/v2/silence/for DELETE (Alertmanager)/api/v2/alerts/groupsfor GET (Alertmanager)/api/v2/alertsfor GET (Alertmanager)
When started with the -enable-label-apis flag, the application can also proxy the following endpoints:
/api/v1/labelsfor GET and POST methods (Prometheus/Thanos)/api/v1/label/<name>/valuesfor GET method (Prometheus/Thanos)
Particularly, you can run prom-label-proxy with label tenant and point to example, demo Prometheus server e.g:
prom-label-proxy \
-label tenant \
-upstream http://demo.do.prometheus.io:9090 \
-insecure-listen-address 127.0.0.1:8080
Accessing demo Prometheus APIs on 127.0.0.1:8080 will now expect tenant query parameter to be set in the URL:
➜ ~ curl http://127.0.0.1:8080/api/v1/query\?query="up"
The "tenant" query parameter must be provided.
➜ ~ curl http://127.0.0.1:8080/api/v1/query\?query="up"\&tenant\="something"
{"status":"success","data":{"resultType":"vector","result":[]}}%You can also provide a static value for a label. For example, running prom-label-proxy with
prom-label-proxy \
-label tenant \
-value prometheus \
-upstream http://demo.do.prometheus.io:9090 \
-insecure-listen-address 127.0.0.1:8080
will enforce tenant=prometheus in all requests.
In this mode, sending the label value as a query parameter will result in the request getting rejected as a 400 Bad Request.
Once again for clarity: this project only enforces a particular label in the respective calls to Prometheus, it in itself does not authenticate or authorize the requesting entity in any way, this has to be built around this project.
The proxy ensures that all selectors passed as matchers to the /federate endpoint must contain that exact match of the particular label (and throws away all other matchers for the label).
For the two query endpoints (/api/v1/query and /api/v1/query_range), the proxy parses the PromQL expression and modifies all selectors in the same way. The label-key is configured as a flag on the binary and the label-value is passed as a query parameter.
For example, if requesting the PromQL query
http_requests_total{namespace=~"a.*"}
and specifying the namespace label must be enforced to b, then the query will be re-written to
http_requests_total{namespace="b"}
This is enforced for any case, whether a label matcher is specified in the original query or not.
Similar to query endpoint, for metadata endpoints /api/v1/series, /api/v1/labels, /api/v1/label/<name>/values the proxy injects the specified label all the provided match[] selectors.
NOTE: When the /api/v1/labels and /api/v1/label/<name>/values endpoints were added to prom-label-proxy, the Prometheus and Thanos endpoints didn't support the match[] parameter hence the prom-label-proxy labels endpoints are disabled by default. Use the -enable-label-apis flag to enable with care. Ensure that the upstream endpoints support label selectors:
The proxy requests the /api/v1/rules Prometheus endpoint, discards the rules that don't contain an exact match of the label and returns the modified response to the client.
The proxy requests the /api/v1/alerts Prometheus endpoint, discards the rules that don't contain an exact match of the label and returns the modified response to the client.
The proxy ensures the following:
GETrequests to the/api/v2/silencesendpoint contain afilterparameter that matches exactly the particular label and throws away all other matchers for the label.POSTrequests to the/api/v2/silencesendpoint can only affect silences that match the label and the label matcher is enforced.DELETErequests to the/api/v2/silence/endpoint can only affect silences that match the label.
The concrete setup being shipped in OpenShift starting with 4.0: the proxy is configured to work with the label-key: namespace. In order to ensure that this is secure is it paired with the kube-rbac-proxy and its URL rewrite functionality, meaning first ServiceAccount token authentication is performed, and then the kube-rbac-proxy authorization to see whether the requesting entity is allowed to retrieve the metrics for the requested namespace. The RBAC role we chose to authorize against is the same as the Kubernetes Resource Metrics API, the reasoning being, if an entity can kubectl top pod in a namespace, it can see cAdvisor metrics (container_memory_rss, container_cpu_usage_seconds_total, etc.).