OPA can periodically report decision logs to remote HTTP servers. 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.

  1. POST /[<decision_logs.resource>] HTTP/1.1
  2. Content-Encoding: gzip
  3. 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.

  1. [
  2. {
  3. "labels": {
  4. "app": "my-example-app",
  5. "id": "1780d507-aea2-45cc-ae50-fa153c8e4a5a",
  6. "version": "v0.32.1"
  7. },
  8. "decision_id": "4ca636c1-55e4-417a-b1d8-4aceb67960d1",
  9. "bundles": {
  10. "authz": {
  11. "revision": "W3sibCI6InN5cy9jYXRhbG9nIiwicyI6NDA3MX1d"
  12. }
  13. },
  14. "path": "http/example/authz/allow",
  15. "input": {
  16. "method": "GET",
  17. "path": "/salary/bob"
  18. },
  19. "result": "true",
  20. "requested_by": "[::1]:59943",
  21. "timestamp": "2018-01-01T00:00:00.000000Z"
  22. }
  23. ]

Decision log updates contain the following fields:

FieldTypeDescription
[].labelsobjectSet of key-value pairs that uniquely identify the OPA instance.
[].decisionidstringUnique 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.

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:

  1. decision_logs:
  2. 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:

  1. {
  2. "resource": "user",
  3. "name": "bob",
  4. "password": "passw0rd"
  5. }

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

  1. package system.log
  2. mask["/input/password"] {
  3. # OPA provides the entire decision log event as input to the masking policy.
  4. # Refer to the original input document under input.input.
  5. input.input.resource == "user"
  6. }
  7. # To mask certain fields unconditionally, omit the rule body.
  8. 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:

  1. {
  2. "decision_id": "b4638167-7fcb-4bc7-9e80-31f5f87cb738",
  3. "erased": [
  4. "/input/password",
  5. "/input/ssn"
  6. ],
  7. "input": {
  8. "name": "bob",
  9. "resource": "user"
  10. },
  11. ------------------------- 8< -------------------------
  12. "path": "system/main",
  13. "requested_by": "127.0.0.1:36412",
  14. "result": true,
  15. "timestamp": "2019-06-03T20:07:16.939402185Z"
  16. }

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

  • Pointers must be prefixed with /input or /result.
  • 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

  1. package system.log
  2. mask[{"op": "upsert", "path": "/input/password", "value": x}] {
  3. # conditionally upsert password if it existed in the orginal event
  4. input.input.password
  5. x := "**REDACTED**"
  6. }

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

  1. package system.log
  2. # always upsert, no conditions in rule body
  3. mask[{"op": "upsert", "path": "/input/password", "value": x}] {
  4. x := "**REDACTED**"
  5. }

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.

  1. {
  2. "decision_id": "b4638167-7fcb-4bc7-9e80-31f5f87cb738",
  3. "erased": [
  4. "/input/ssn"
  5. ],
  6. "masked": [
  7. "/input/password"
  8. ],
  9. "input": {
  10. "name": "bob",
  11. "resource": "user",
  12. "password": "**REDACTED**"
  13. },
  14. ------------------------- 8< -------------------------
  15. "path": "system/main",
  16. "requested_by": "127.0.0.1:36412",
  17. "result": true,
  18. "timestamp": "2019-06-03T20:07:16.939402185Z"
  19. }