[
https://issues.apache.org/jira/browse/RANGER-5655?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ramachandran Krishnan updated RANGER-5655:
------------------------------------------
Attachment: (was: Dynamic Ingestor Registry Guide (Ranger Audit
Ingestor).pdf)
> Implement dynamic unified ingestor registry for audit-ingestor: runtime Kafka
> partition routing and per-repo service allowlists via compacted topic + REST,
> without ingestor restarts. Feature flag default off.
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: RANGER-5655
> URL: https://issues.apache.org/jira/browse/RANGER-5655
> Project: Ranger
> Issue Type: Improvement
> Components: Ranger
> Reporter: Ramachandran Krishnan
> Assignee: Ramachandran Krishnan
> Priority: Major
> Fix For: 3.0.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Implement a dynamic unified ingestor registry for Ranger audit-ingestor so
> operators can change Kafka partition routing and per-repo service allowlists
> at runtime — without restarting ingestor pods.
> The registry is stored in a Kafka compacted topic
> ({{{}ranger_audit_partition_plan{}}}) and managed via REST
> ({{{}/api/audit/partition-plan/*{}}}). All ingestor replicas converge on the
> same versioned plan through a background watcher; {{AuditPartitioner}} routes
> audit records on the hot path from in-memory state only.
> Feature flag (default off):
> {{ranger.audit.ingestor.kafka.partition.plan.dynamic.enabled=false}}
> ----
> h2. Problem
>
> Today (static mode), audit-ingestor loads two kinds of configuration from XML
> at startup only:
> ||Job||Question||Static behavior||
> |Service allowlist|May this Kerberos principal POST audits for repo
> R?|{{ranger.audit.ingestor.service.<repo>.allowed.users}} in site XML|
> |Partition routing|After accept, which {{ranger_audits}}
> partition?|{{kafka.configured.plugins}} + per-plugin overrides in site XML|
> Changing either requires editing XML and restarting every ingestor replica.
> Contiguous-range static allocation can also reshuffle later plugins when an
> early plugin's partition count changes.
> *Goals:*
> * Onboard new plugins/repos and scale hot plugins without ingestor restart
> * Append-only partition growth (no reshuffle of existing plugin assignments)
> * One shared source of truth across all ingestor pods
> * No new infra (no Postgres / ZooKeeper for the registry)
> ----
> h2. Solution
> Introduce a unified partition plan document (versioned JSON) in Kafka topic
> {{ranger_audit_partition_plan}} (1 partition, compacted). One document holds:
> * {{plugins}} — dedicated partition IDs per plugin id (Kafka record key /
> agent id)
> * {{buffer}} — partition pool for not-yet-promoted plugins (sticky hash)
> * {{services}} — per-repo {{allowedUsers}} for POST {{/api/audit/access}}
> (optional {{pluginId}} for ownership)
> * {{topicPartitionCount}} — must match live {{ranger_audits}} partition count
> * {{version}} — optimistic locking for REST mutations
> *Control plane:* REST +registry topic+ {{PartitionPlanWatcher}}
> *Data plane:* plugins POST {{/api/audit/access}} → allowlist check →
> {{AuditPartitioner}} → {{ranger_audits}}
> Solr/HDFS dispatchers are unchanged; they consume all partitions of
> {{{}ranger_audits{}}}.
>
> h2. REST API (3 endpoints)
> All mutations use optimistic locking via {{expectedVersion}} (from
> {{{}GET{}}}). Stale writes return *409* with the current plan body.
> ||Method|| ||Endpoint||Purpose||
> |{{GET}}| |{{/api/audit/partition-plan}}|Read current plan (503 when dynamic
> mode off)|
> |{{POST}}| |{{/api/audit/partition-plan/plugins}}|*Onboard* plugin: promote
> from buffer + register service allowlists in one version bump|
> |{{PATCH}}| |{{{}/api/audit/partition-plan/plugins/{pluginId{}}}}|*Update*
> onboarded plugin: scale partitions and/or mutate allowlists
> |
> | | | | |
> *{{removeServices}} semantics:* removes the whole {{services[repo]}} entry
> (including {{allowedUsers}} and {{{}pluginId{}}}). Does *not* change
> {{{}plugins.{*}.partitions{*}{}}}{*}. On each ingestor pod,
> {{AuthToLocalRuleComposer}} recomposes {{auth_to_local}} from the *union of
> remaining{*} {{{}services[*].allowedUsers{}}}; short names only on removed
> repos no longer map → *403* on {{{}/access{}}}.
> *Future work (not in this change):* bootstrap *v0* split — seeding
> {{services}} from XML separately from plugin partition assignments. Current
> bootstrap still publishes a single v1 plan from XML.
> ----
> h2. Key deliverables
> * Kafka registry — {{{}KafkaPartitionPlanRegistry{}}}, bootstrap from XML on
> greenfield, brownfield pre-seed support
> * REST API — *3 endpoints* above with {{expectedVersion}} (409 on stale)
> * Dynamic partitioner — {{AuditPartitioner}} reads
> {{{}PartitionPlanHolder{}}}; round-robin for promoted plugins; buffer sticky
> hash for unknown plugins; post-scale routing uses max(cluster, plan) when
> metadata lags
> * Unified allowlist — {{services}} map in same registry; {{auth_to_local}}
> rules recomposed from allowlist union in dynamic mode; per-repo POST checks
> use each repo's own allowlist
> * Topic grow — grow {{ranger_audits}} before registry write when
> onboarding/scaling
> * E2E harness — Docker Tier 3 scripts for partition-plan REST, plugin
> onboard/routing, {{{}auth_to_local{}}}, full plugin→Solr pipelines (HDFS,
> Ozone, Hive, etc.)
> * Documentation — implementation guide, ops runbook, brownfield migration,
> E2E test plan
>
>
> ----
> h2. Configuration
> {code:java}
> <property>
> <name>ranger.audit.ingestor.kafka.partition.plan.dynamic.enabled</name>
> <value>true</value>
> </property>
> <property>
> <name>ranger.audit.ingestor.kafka.partition.plan.topic</name>
> <value>ranger_audit_partition_plan</value>
> </property> {code}
> When dynamic mode is on and registry is empty, the first ingestor pod
> bootstraps an initial plan from existing XML properties
> ({{{}kafka.configured.plugins{}}}, buffer/per-plugin counts, service
> allowlists).
>
> h2. Example plan JSON
> {code:java}
> {
> "topic": "ranger_audits",
> "version": 12,
> "topicPartitionCount": 48,
> "plugins": {
> "hdfs": { "partitions": [0, 1, 2, 3, 4, 5] },
> "hiveServer2": { "partitions": [6, 7, 8, 9, 10, 11] }
> },
> "buffer": { "partitions": [12, 13, "..."] },
> "services": {
> "dev_hive": { "allowedUsers": ["hive"], "pluginId": "hiveServer2" },
> "dev_ozone": { "allowedUsers": ["om", "ozone"], "pluginId": "ozone" },
> "dev_hdfs": { "allowedUsers": ["hdfs", "nn"], "pluginId": "hdfs" }
> }
> }{code}
> h2. Testing
> h3. Unit tests and quality gates
> export MAVEN_OPTS="-Xmx4g"
> mvn verify -pl audit-server/audit-common,audit-server/audit-ingestor
> -Drat.skip=true}}
> ||Gate||Result||
> |Unit tests|*94 passed* — {{audit-common}} + {{audit-ingestor}}|
> |Checkstyle|*0 violations*|
> |PMD|*0 rule violations*|
> h3. Validated scenarios
> h4. 1. Static vs dynamic mode
> What it proves: The feature flag works and existing deployments are not
> broken.
> ||Mode||Config||Expected behavior||
> |Static (default)|{{kafka.partition.plan.dynamic.enabled=false}}|Routing and
> allowlists from XML at startup only. {{GET /api/audit/partition-plan}}
> returns {*}503{*}. Ingestor health {*}200{*}. Plugin audits still flow to
> Solr.|
> |Dynamic|{{dynamic.enabled=true}}|{{PartitionPlanWatcher}} active; routing
> from in-memory plan. {{GET /partition-plan}} returns JSON.|
> h4. 2. Greenfield bootstrap
> What it proves: On a new cluster with an empty plan topic, the first ingestor
> pod creates version 1 of the partition plan automatically — no manual REST
> call.
> Flow:
> # Enable dynamic mode; restart ingestor.
> # Plan topic {{ranger_audit_partition_plan}} is created (1 partition,
> compacted).
> # Ingestor publishes v1 from existing XML ({{{}kafka.configured.plugins{}}},
> buffer size, per-plugin overrides, service allowlists).
> # Plan's {{topicPartitionCount}} matches live {{ranger_audits}} partition
> count (e.g. 48 = 13 plugins × 3 + 9 buffer on full lab layout).
> Success: {{GET /partition-plan}} shows v1, correct plugin list, partition
> count aligned with Kafka. Audits still work after enable.
> h4. 3. REST onboard / update (200 / 409 / 400)
> What it proves: Operators can change routing and allowlists at runtime via
> REST without restarting ingestor. The API rejects bad requests.
> ||Action||REST||Success||Failure cases||
> |Onboard plugin from buffer|{{POST /api/audit/partition-plan/plugins}} (with
> non-empty {{{}services{}}})|200, version increments, plugin in {{plugins}}
> map, services merged|400 if plugin already promoted; 400 if {{services}}
> missing/empty|
> |Scale hot plugin (+N partitions)|{{{}PATCH .../plugins/{pluginId{}}}} with
> {{additionalPartitions}}|200, new partition IDs appended at tail;
> {{ranger_audits}} grown first if needed|400 if scaling buffer-only / unknown
> plugin|
> |Mutate allowlists|{{{}PATCH .../plugins/{pluginId{}}}} with {{addServices}}
> / {{updateServices}} / {{removeServices}}|200, {{services}} updated;
> {{auth_to_local}} recomposed on plan install|400 if repo owned by another
> plugin|
> Key property: Changes are append-only — existing plugin partition lists are
> never reshuffled.
> h4. 4. Multi-pod convergence
> What it proves: In a real cluster with multiple ingestor replicas behind a
> load balancer, all pods see the same plan after an admin change.
> Flow:
> # Primary ingestor on :7081, second replica on :7082 (same Kafka plan topic,
> different Kerberos identity).
> # Both pods report the same version at startup.
> # Admin promotes or updates a plugin on primary only.
> # Within one watcher cycle (~30s), replica on :7082 shows the same new
> version without any REST call on the replica.
> Success: No drift between pods; routing is consistent cluster-wide.
> h4. 5. Brownfield pre-seed
> What it proves: Existing production clusters can cut over to dynamic mode
> safely by writing the plan into Kafka before enabling the feature — ingestor
> must not overwrite it with a fresh XML bootstrap.
> Flow (Path A migration):
> # Capture current plan JSON while dynamic is briefly on.
> # Turn dynamic off; delete plan topic (simulate “still on static”).
> # Operator pre-seeds plan to Kafka with {{version=1}} and marker
> {{{}updatedBy=brownfield-e2e-seed{}}}.
> # Enable dynamic; restart ingestor.
> # Ingestor adopts pre-seeded plan — not auto-bootstrap from XML.
> # Rollback: turn dynamic off → partition-plan API {*}503{*}, health
> {*}200{*}, audits still OK in static mode.
> h4. 6. Plugin onboard + Kafka partition routing
> What it proves: A real plugin can be onboarded via REST and its audit events
> land on the correct Kafka partitions defined in the plan.
> Flow:
> # Enable dynamic mode (greenfield buffer-only layout).
> # For each running plugin container (HDFS, Ozone, Hive, etc.): {{POST
> /api/audit/partition-plan/plugins}} with {{{}pluginId{}}},
> {{{}partitionCount{}}}, {{{}services{}}}.
> # Plugin authenticates with Kerberos and POST {{{}/api/audit/access{}}}.
> # Read Kafka record for that event → partition number must be in the
> plugin's assigned list in the plan (not buffer, after onboard).
> h4. 7. auth_to_local recomposition
> What it proves: The unified registry {{services}} map controls who may POST
> audits, and Kerberos principals are mapped to short usernames correctly.
> Flow (per plugin repo):
> # Plugin calls POST {{/access}} with Kerberos principal (e.g.
> {{{}hdfs/[email protected]{}}}) → ingestor maps via
> {{auth_to_local}} → short name {{hdfs}} → *200* if in
> {{{}services[dev_hdfs].allowedUsers{}}}.
> # Remove user from allowlist via {{{}PATCH .../plugins/{pluginId{}}}}
> ({{{}updateServices{}}} or {{{}removeServices{}}}) → same principal gets
> {*}403{*}; {{removeServices}} drops the whole repo entry and recomposes
> {{auth_to_local}} from remaining allowlists.
> # Re-add allowlist → *200* again.
> # Cross-repo denial: HDFS principal posting to {{dev_kms}} repo → {*}403{*}.
> Why it matters: In dynamic mode, allowlists live in the registry (not XML
> restart). {{auth_to_local}} rules must stay in sync with
> {{{}services[].allowedUsers{}}}.
> h4. 8. HDFS / Ozone / Hive full audit pipelines
> What it proves: Dynamic partition plan changes do not break the end-to-end
> audit path: plugin → ingestor → Kafka → dispatcher → Solr (and Admin Audit UI
> when {{{}audit_store=solr{}}}).
> ||Plugin||What is exercised||
> |HDFS|Real NameNode operation → plugin audit → ingestor → correct partition →
> Solr doc|
> |Ozone|OM principal ({{{}om{}}}, {{{}ozone{}}}) → {{dev_ozone}} repo → full
> pipeline|
> |Hive|HS2 principal ({{{}hive{}}}) → {{dev_hive}} repo → full pipeline|
> Validated after onboard/scale (partition plan changed live), not only at
> bootstrap.
> Verify Solr:
>
> {{curl -s
> '[http://localhost:8983/solr/ranger_audits/select?q=repo:dev_hdfs&rows=3&wt=json'|http://localhost:8983/solr/ranger_audits/select?q=repo:dev_hdfs&rows=3&wt=json%27]}}
> h3. Summary table
> ||Scenario||Plain question||How verified||Pass criteria||
> |Static vs dynamic|Does default-off behavior still work?|Feature flag + REST
> 503/200|Static unchanged; dynamic enables plan API|
> |Greenfield bootstrap|Who creates v1 on empty topic?|First ingestor
> start|Plan v1 in Kafka; counts match|
> |REST onboard/update|Can ops change routing + allowlists live?|REST
> mutations|200 on valid; 409 stale; 400 illegal|
> |Multi-pod convergence|Do all replicas agree?|2 ingestors, update on one|Same
> version ≤35s on both|
> |Brownfield pre-seed|Safe production cutover?|Pre-write plan, then
> enable|Pre-seed preserved; rollback OK|
> |Plugin onboard + routing|Do audits hit right partitions?|POST access + Kafka
> inspect|Partition ∈ plan list|
> |auth_to_local|Does allowlist enforcement work?|Allow/deny via services
> map|200/403 per principal+repo; recomposition on remove|
> |Full pipelines|Is audit delivery intact?|HDFS/Ozone/Hive → Solr|Docs in
> {{ranger_audits}} Solr collection
>
> |
>
> ----
> h2. Acceptance criteria
> # With {{{}dynamic.enabled=false{}}}, behavior matches existing static XML
> partitioning; {{GET /partition-plan}} returns *503*
> # With {{{}dynamic.enabled=true{}}}, registry topic created (1 partition,
> compacted); bootstrap plan published on greenfield
> # {{POST .../plugins}} onboards plugin + services; {{{}PATCH
> .../plugins/{pluginId{}}}} scales and/or mutates allowlists; stale
> {{expectedVersion}} returns *409*
> # All ingestor replicas converge to same plan within watcher refresh interval
> # Audits for promoted plugin land only on assigned {{ranger_audits}}
> partitions
> # Service allowlists updated via {{{}PATCH .../plugins/{pluginId{}}}}
> ({{{}addServices{}}} / {{updateServices}} / {{{}removeServices{}}});
> unauthorized principal returns {*}403{*}; {{removeServices}} recomposes
> {{auth_to_local}} from remaining allowlists
> # Post-scale routing works when Kafka metadata lags plan
> ({{{}AuditPartitioner{}}} bound logic)
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
