graphql-monitoring-proxy

graphql monitoring proxy

Creates a passthrough proxy to a graphql endpoint(s), allowing you to analyse the queries and responses, producing the Prometheus metrics at a fraction of the cost - because, as we know - $0 is a fair price.

This project is in active use by telegram-bot.app, and was tested with 30k queries per second on a single instance, consuming 10 MB of RAM and 0.1% CPU.

• graphql monitoring proxy
  • Why this project exists
  • How to deploy
    • Note on websocket support
  • Endpoints
  • Features
  • Configuration
  • Speed
    • Caching
    • Read-only endpoint
  • Security
    • Role-based rate limiting
    • Read-only mode
    • Allowing access to listed URLs
    • Blocking introspection
  • API endpoints
    • Ban or unban the user
  • General
    • Metrics which matter
    • Healthcheck
    • Monitoring endpoint

Why this project exists

I wanted to monitor the queries and responses of our graphql endpoint. Still, we didn't want to pay the price of the graphql server itself ( and I will not point fingers at a particular well-known project), as monitoring and basic security features should be a standard, free functionality.

How to deploy

You can find the example of the Kubernetes manifest in the example standalone deployment or example combined deployment files. Observed advantage of multideployment is that it allows the network requests to travel via localhost, without leaving the deployment which brings quite significant network performance boost.

Note on websocket support

Proxy in its current version 0.5.30 does not support websockets. If you need to proxy the websocket requests - you can use following trick whilst setting up the proxy. As I'm a big fan of Traefik - there's an example which works with the mentioned above combined deployment.

apiVersion: traefik.containo.us/v1alpha1
kind: IngressRoute
metadata:
  name: hasura-internal
spec:
  entryPoints:
    - websecure
  routes:
    # NON WEBSOCKET CONNECTION
    - kind: Rule
      match: Host(`example.com`) && PathPrefix(`/v1/graphql`) && !HeadersRegexp(`Upgrade`, `websocket`)
      services:
      - name: hasura-w-proxy-internal
        port: proxy
      middlewares:
        - name: compression
          namespace: default

    # WEBSOCKET CONNECTION
    - kind: Rule
      match: Host(`example.com`) && PathPrefix(`/v1/graphql`) && HeadersRegexp(`Upgrade`, `websocket`)
      services:
      - name: hasura-w-proxy-internal
        port: hasura
      middlewares:
        - name: compression
          namespace: default

Endpoints

• :8080/* - the graphql passthrough endpoint
• :9393/metrics - the prometheus metrics endpoint
• :8080/healthz - the healthcheck endpoint
• :8080/livez - the liveness probe endpoint
• :9090/api/* - the monitoring proxy API endpoint

Features

Configuration

All the environment variables should be prefixed with GMP_ to avoid conflicts with other applications. If GMP_ prefixed environment variable is present - it will take precedence over the non-prefixed one. You can still use the non-prefixed environment variables in the spirit of the backward compatibility, but it's not recommended.

Speed

Caching

The cache engine is enabled in the background by default, using no additional resources. You can then start using the cache by setting the ENABLE_GLOBAL_CACHE environment variable to true - which will enable the cache for all queries without introspection. You can leave the global cache disabled and enable the cache for specific queries by adding the @cached directive to the query.

In the case of the @cached you can add additional parameters to the directive which will set the cache for specific queries to the provided time. For example, query MyCachedQuery @cached(ttl: 90) .... will set the cache for the query to 90 seconds.

You can also set cache for specific query by using X-Cache-Graphql-Query header, which will set the cache for the query to the provided time, for example X-Cache-Graphql-Query: 90 will set the cache for the query to 90 seconds.

You can also force refresh of the cache by using @cached(refresh: true) directive in the query, for example:

query MyProducts @cached(refresh: true) {
  products {
    id
    name
  }
}

Since version 0.5.30 the cache is gzipped in the memory, which should optimise the memory usage quite significantly.

Read-only endpoint

You can now specify the read-only GraphQL endpoint by setting the HOST_GRAPHQL_READONLY environment variable. The default value is empty, preventing the proxy from using the read-only endpoint for the queries and directing all the requests to the main endpoint specified as HOST_GRAPHQL. If the HOST_GRAPHQL_READONLY is set, the proxy will use the read-only endpoint for the queries with the query type and the main endpoint for the mutation type queries. Format of the read-only endpoint is the same as HOST_GRAPHQL endpoint, for example http://localhost:8080/.

You can check out the example of combined deployment with RW and read-only hasura.

Security

Role-based rate limiting

You can rate limit requests using the ROLE_RATE_LIMIT environment variable. If enabled, the proxy will rate limit the requests based on the role claim in the JWT token. You can then provide the JSON file in the following format to specify the limits. The default interval is second, but you can use other values as well. If you want to disable the rate limiting for a specific role, you can set the req to 0.

Available values: nano, micro, milli, second, minute, hour, day

To define path in JWT token where the current user role is present, use the JWT_ROLE_CLAIM_PATH environment variable.

You can also set up the ROLE_FROM_HEADER environment variable to extract the role from the header instead of the JWT token. This is useful if you want to rate limit the requests for unauthenticated users. It's worth mentioning that ROLE_FROM_HEADER takes a priority over the JWT_ROLE_CLAIM_PATH environment variable and if its set, the proxy will not try to extract the role from the JWT token.

Default/sample configuration:

{
  "ratelimit": {
      "admin": {
          "req": 100,
          "interval": "second"
      },
      "guest": {
          "req": 50,
          "interval": "minute"
      },
      "-": {
          "req": 100,
          "interval": "day"
      }
  }
}

If you'd like to change it - mount your configmap as /app/ratelimit.json file. Remember to include the - role, which is used for unauthenticated users or when claim can't be found for any reason. If rate limit has been reached - the proxy will return 429 Too Many Requests error.

Read-only mode

You can enable the read-only mode by setting the READ_ONLY_MODE environment variable to true - which will block all the mutation queries.

Allowing access to listed URLs

You can allow access only to certain URLs by setting the ALLOWED_URLS environment variable to a comma-separated list of URLs. If enabled - other URLs will return 403 Forbidden error and request will not reach the proxied service.

Blocking introspection

You can block the schema introspection by setting the BLOCK_SCHEMA_INTROSPECTION environment variable to true - which will block all the queries with introspection parts, like:

__schema, __type, __typename, __directive, __directivelocation, __field, __inputvalue, __enumvalue, __typekind, __fieldtype, __inputobjecttype, __enumtype, __uniontype, __scalars, __objects, __interfaces, __unions, __enums, __inputobjects, __directives

If you'd like to keep blocking of the schema introspection on but allow one or more of from the list of above for any reason, you can use the ALLOWED_INTROSPECTION environment variable to specify the list of allowed queries.

ALLOWED_INTROSPECTION="__typename,__type"

API endpoints

Ban or unban the user

Your monitoring system can detect user misbehaving, for example trying to extract / scrap the data. To prevent user from doing so you can use the simple API to ban the user from accessing the application.

To do so - you need to enable the api by setting env variable ENABLE_API=true which will expose the API on the port API_PORT=9090. Nedless to say - keep it secure and don't expose it outside of your cluster.

Then you can use the following endpoints:

• POST /api/user-ban - ban the user from accessing the application
• POST /api/user-unban - unban the user from accessing the application

Both endpoints require the user_id parameter to be present in the request body and allow you to provide the reason for the ban.

Example request:

curl -X POST \
  http://localhost:9090/api/user-ban \
  -H 'Content-Type: application/json' \
  -d '{
      "user_id": "1337",
      "reason": "Scraping data"
    }'

Ban details will be stored in the banned_users.json file, which you can mount as a file or configmap to the /go/src/app/banned_users.json path ( or use BANNED_USERS_FILE environment variable to specify the path to the file). The file operation is important if you have multiple instances of the proxy running, as it will allow you to ban the user from accessing the application on all instances.

General

Metrics which matter

You can always enable PURGE_METRICS_ON_CRAWL environment variable to purge the metrics on each /metrics crawl. This will allow you to see only the current metrics, without potential leftovers from the previous crawls. This is useful if you want to monitor the metrics in real-time and / or limit the amount of data ingested into the monitoring system. When enabled you will most likely need to update your monitoring queries.

With the PURGE_METRICS_ON_CRAWL enabled, the graphql_proxy_requests_failed, graphql_proxy_requests_skipped and graphql_proxy_requests_succesful metrics will remain between resets.

If you prefer more control over the metrics purging - you can enable PURGE_METRICS_ON_TIMER environment variable and set the interval in seconds. This will allow you to purge the metrics on a regular basis, for example every 90 seconds. It could be better solution if you have multiple crawlers checking the metrics endpoints and you want to avoid the situation when metrics are purged by for example healthcheck.

Healthcheck

If you'd like the /healthz endpoint to perform actual check for the connectivity to the graphql endpoint - set the HEALTHCHECK_GRAPHQL_URL environment variable to the exact URL of the graphql endpoint. The query executed will be query { __typename } and if the response is not 200 OK - the healthcheck will fail. Remember that the endpoint is a full URL which you'd like to check, so it should include the protocol, host and path - for example http://localhost:8080/v1/graphql and it's NOT the same as value of HOST_GRAPHQL environment variable which should provide only the host, without path, ending with slash.

Monitoring endpoint

Example metrics produced by the proxy:

graphql_proxy_timed_query_bucket{cached="false",user_id="-",op_type="mutation",op_name="updateUserDetails",vmrange="1.000e-02...1.136e-02"} 6
graphql_proxy_timed_query_count{op_name="",cached="false",user_id="-",op_type=""} 78
graphql_proxy_timed_query_bucket{op_name="MyQuery",cached="false",user_id="-",op_type="query",vmrange="5.995e+00...6.813e+00"} 1
graphql_proxy_timed_query_sum{op_name="MyQuery",cached="false",user_id="-",op_type="query"} 6
graphql_proxy_timed_query_count{op_name="MyQuery",cached="false",user_id="-",op_type="query"} 1
graphql_proxy_executed_query{user_id="-",op_type="mutation",op_name="updateKnownSpammer",cached="false"} 1486
graphql_proxy_executed_query{user_id="-",op_type="query",op_name="checkIfAdminsNeedRefreshing",cached="false"} 13167
graphql_proxy_executed_query{user_id="1337",op_type="query",op_name="checkIfKnownMedia",cached="false"} 429
graphql_proxy_executed_query{user_id="-",op_type="query",op_name="checkIfSpamAIRequiresUpdate",cached="false"} 8891
graphql_proxy_requests_failed 324
graphql_proxy_requests_skipped 0
graphql_proxy_requests_succesful 454823