Decision Logs Edit

OPA can periodically report decision logs to remote HTTP servers, using custom plugins, or to the console output; or any combination thereof. The decision logs contain events that describe policy queries. Each event includes the policy that was queried, the input to the query, bundle metadata, and other information that enables auditing and offline debugging of policy decisions.

When decision logging is enabled the OPA server will include a decision_id field in API calls that return policy decisions.

See the Configuration Reference for configuration details.

Decision Log Service API

OPA expects the service to expose an API endpoint that will receive decision logs.

POST /[<decision_logs.resource>] HTTP/1.1
Content-Encoding: gzip
Content-Type: application/json

The resource field is an optional configuration that can be used to route logs to a specific endpoint in the service by defining the full path. If the resource path is not configured on the agent, updates will be sent to /logs.

The message body contains a gzip compressed JSON array. Each array element (event) represents a policy decision returned by OPA.

[
  {
    "labels": {
      "app": "my-example-app",
      "id": "1780d507-aea2-45cc-ae50-fa153c8e4a5a",
      "version": "v0.45.0"
    },
    "decision_id": "4ca636c1-55e4-417a-b1d8-4aceb67960d1",
    "bundles": {
      "authz": {
        "revision": "W3sibCI6InN5cy9jYXRhbG9nIiwicyI6NDA3MX1d"
      }
    },
    "path": "http/example/authz/allow",
    "input": {
      "method": "GET",
      "path": "/salary/bob"
    },
    "result": "true",
    "requested_by": "[::1]:59943",
    "timestamp": "2018-01-01T00:00:00.000000Z"
  }
]

Decision log updates contain the following fields:

FieldTypeDescription
[_].labelsobjectSet of key-value pairs that uniquely identify the OPA instance.
[_].decision_idstringUnique identifier generated for each decision for traceability.
[_].bundlesobjectSet of key-value pairs describing the bundles which contained policy used to produce the decision.
[_].bundles[_].revisionstringRevision of the bundle at the time of evaluation.
[_].pathstringHierarchical policy decision path, e.g., /http/example/authz/allow. Receivers should tolerate slash-prefixed paths.
[_].querystringAd-hoc Rego query received by Query API.
[_].inputanyInput data provided in the policy query.
[_].resultanyPolicy decision returned to the client, e.g., true or false.
[_].requested_bystringIdentifier for client that executed policy query, e.g., the client address.
[_].timestampstringRFC3999 timestamp of policy decision.
[_].metricsobjectKey-value pairs of performance metrics.
[_].erasedarray[string]Set of JSON Pointers specifying fields in the event that were erased.
[_].maskedarray[string]Set of JSON Pointers specifying fields in the event that were masked.
[_].nd_builtin_cacheobjectKey-value pairs of non-deterministic builtin names, paired with objects specifying the input/output mappings for each unique invocation of that builtin during policy evaluation. Intended for use in debugging and decision replay. Receivers will need to decode the JSON using Rego’s JSON decoders.

If the decision log was successfully uploaded to the remote service, it should respond with an HTTP 2xx status. If the service responds with a non-2xx status, OPA will requeue the last chunk containing decision log events and upload it during the next upload event. OPA also performs an exponential backoff to calculate the delay in uploading the next chunk when the remote service responds with a non-2xx status.

OPA periodically uploads decision logs to the remote service. In order to conserve network and memory resources, OPA attempts to fill up each upload chunk with as many events as possible while respecting the user-specified upload_size_limit_bytes config option. OPA defines an adaptive (soft) limit that acts as a measure for encoding as many decisions into each chunk as possible. It uses the below algorithm to optimize the number of log events to include in a chunk. The algorithm features three phases namely:

Scale Up: If the current chunk size is within 90% of the user-configured (hard) limit, exponentially increase the soft limit. The exponential function is 2^x where x has a minimum value of 1

Scale Down: If the current chunk size exceeds the hard limit, decrease the soft limit and re-encode the decisions in the last chunk.

Equilibrium: If the chunk size is between 90% and 100% of the user-configured limit, maintain soft limit value.

When an event containing nd_builtin_cache cannot fit into a chunk smaller than upload_size_limit_bytes, OPA will drop the nd_builtin_cache key from the event, and will retry encoding the chunk without the non-deterministic builtins cache information. This best-effort approach ensures that OPA reports decision log events as much as possible, and bounds how large decision log events can get. This size-bounding is necessary, because some non-deterministic builtins (such as http.send) can increase the decision log event size by a potentially unbounded amount.

Local Decision Logs

Local console logging of decisions can be enabled via the console config option. This does not require any remote server. Example of minimal config to enable:

decision_logs:
    console: true

This will dump all decisions to the console. See Configuration Reference for more details.

Masking Sensitive Data

Policy queries may contain sensitive information in the input document that must be removed or modified before decision logs are uploaded to the remote API (e.g., usernames, passwords, etc.) Similarly, parts of the policy decision itself may be considered sensitive.

By default, OPA queries the data.system.log.mask path prior to encoding and uploading decision logs or calling custom decision log plugins.

OPA provides the decision log event as input to the policy query and expects the query to return a set of JSON Pointers that refer to fields in the decision log event to either erase or modify.

For example, assume OPA is queried with the following input document:

{
  "resource": "user",
  "name": "bob",
  "password": "passw0rd"
}

To remove the password field from decision log events related to “user” resources, supply the following policy to OPA:

package system.log

mask["/input/password"] {
  # OPA provides the entire decision log event as input to the masking policy.
  # Refer to the original input document under input.input.
  input.input.resource == "user"
}

# To mask certain fields unconditionally, omit the rule body.
mask["/input/ssn"]

When the masking policy generates one or more JSON Pointers, they will be erased from the decision log event. The erased paths are recorded on the event itself:

{
  "decision_id": "b4638167-7fcb-4bc7-9e80-31f5f87cb738",
  "erased": [
    "/input/password",
    "/input/ssn"
  ],
  "input": {
    "name": "bob",
    "resource": "user"
  },
------------------------- 8< -------------------------
  "path": "system/main",
  "requested_by": "127.0.0.1:36412",
  "result": true,
  "timestamp": "2019-06-03T20:07:16.939402185Z"
}

There are a few restrictions on the JSON Pointers that OPA will erase:

  • Pointers must be prefixed with /input, /result, or /nd_builtin_cache.
  • Pointers may be undefined. For example /input/name/first in the example above would be undefined. Undefined pointers are ignored.
  • Pointers must refer to object keys. Pointers to array elements will be treated as undefined. For example /input/emails/0/value is allowed but /input/emails/0 is not.

In order to modify the contents of an input field, the mask rule may utilize the following format.

  • "op" – The operation to apply when masking. All operations are done at the path specified. Valid options include:
opDescription
"remove"The "path" specified will be removed from the resulting log message. The "value" mask field is ignored for "remove" operations.
"upsert"The "value" will be set at the specified "path". If the field exists it is overwritten, if it does not exist it will be added to the resulting log message.
  • "path" – A JSON pointer path to the field to perform the operation on.

Optional Fields:

  • "value" – Only required for "upsert" operations.

This is processed for every decision being logged, so be mindful of performance when performing complex operations in the mask body, eg. crypto operations

package system.log

mask[{"op": "upsert", "path": "/input/password", "value": x}] {
  # conditionally upsert password if it existed in the orginal event
  input.input.password
  x := "**REDACTED**"
}

To always upsert a value, even if it didn’t exist in the original event, the following rule format can be used.

package system.log

# always upsert, no conditions in rule body
mask[{"op": "upsert", "path": "/input/password", "value": x}] {
  x := "**REDACTED**"
}

The result of this mask operation on the decision log event produces the following output. Notice that the mask event field exists to track remove vs upsert mask operations.

{
  "decision_id": "b4638167-7fcb-4bc7-9e80-31f5f87cb738",
  "erased": [
    "/input/ssn"
  ],
  "masked": [
    "/input/password"
  ],
  "input": {
    "name": "bob",
    "resource": "user",
    "password": "**REDACTED**"
  },
------------------------- 8< -------------------------
  "path": "system/main",
  "requested_by": "127.0.0.1:36412",
  "result": true,
  "timestamp": "2019-06-03T20:07:16.939402185Z"
}

Rate Limiting Decision Logs

There are scenarios where OPA may be uploading decisions faster than what the remote service is able to consume. Although OPA provides a user-specified buffer size limit in bytes, it may be difficult to determine the ideal buffer size that will allow the service to consume logs without being overwhelmed. The max_decisions_per_second config option allows users to set the maximum number of decision log events to buffer per second. OPA will drop events if the rate limit is exceeded. This option provides users more control over how OPA buffers log events and is an effective mechanism to make sure the service can successfully process incoming log events.