jerryshao commented on code in PR #10696:
URL: https://github.com/apache/gravitino/pull/10696#discussion_r3078654598
##########
design-docs/cache-improvement-design.md:
##########
@@ -0,0 +1,873 @@
+# Gravitino Cache Improvement Design
+
+---
+
+## 1. Background
+
+### 1.1 System Overview
+
+Gravitino is a unified metadata management control plane. Compute engines
(Spark, Flink, Trino)
+call it during query planning to resolve catalog, schema, and table metadata,
and to verify
+user permissions. The access pattern is distinctly **read-heavy,
write-light**: DDL operations
+are infrequent, and metadata is resolved once per job.
+
+Gravitino is evolving from single-node to multi-node active-active HA
deployment. Each node
+currently maintains its own independent in-process Caffeine cache with no
cross-node
+synchronisation. Under HA, any write on one node leaves other nodes' caches
stale until TTL
+expiry.
+
+---
+
+### 1.2 Current Cache Architecture Overview
+
+Gravitino maintains three distinct caching layers for the authorization path:
+
+```
+┌──────────────────────────────────────────────────────┐
+│ Layer 3: Per-request cache (AuthorizationRequestContext) │
+│ Scope: one HTTP request; prevents duplicate auth calls │
+├──────────────────────────────────────────────────────┤
+│ Layer 2: Auth policy caches (JcasbinAuthorizer) │
+│ loadedRoles Cache<Long, Boolean> hook update/TTL │
+│ ownerRel Cache<Long, Optional<Long>> hook update/TTL │
+├──────────────────────────────────────────────────────┤
+│ Layer 1: Entity store cache (RelationalEntityStore) │
+│ CaffeineEntityCache — or NoOpsCache when disabled │
+│ Caches entity reads and relation queries for all modules │
+│ Controlled by Configs.CACHE_ENABLED │
+└──────────────────────────────────────────────────────┘
+```
+
+**JCasbin is the core of the auth cache system.** It maintains an in-memory
policy table:
+
+```
+(roleId, objectType, metadataId, privilege) → ALLOW | DENY
+```
+
+The Layer 2 caches exist solely to manage JCasbin's policy loading lifecycle:
+
+| Cache | Role
|
+|-----------------------------------------|----------------------------------------------------------------------------------------------------------------|
+| `loadedRoles: Cache<Long, Boolean>` | Tracks which roles are already
loaded into JCasbin — prevents repeated [C2]+[C3] queries on every auth request
|
+| `ownerRel: Cache<Long, Optional<Long>>` | Caches owner lookups — **prevents
[D1] on every auth request** (2–4 `isOwner()` calls per request, see §1.3.2) |
+
+Without `loadedRoles`, every auth request would re-execute N DB queries to
reload all of a
+user's role policies into JCasbin. These two caches are the reason the auth
path is fast on
+the warm path. Layer 1 (entity cache) additionally accelerates the name→ID
resolution calls
+([A], [B], [C1]) that feed into JCasbin's enforce call.
+
+---
+
+#### 1.2.1 Problems with the Current Entity Cache
+
+**The entity cache (Layer 1) has accumulated significant complexity and is not
well-suited to
+serve as a general-purpose or auth-dedicated caching layer.**
+
+##### Mixed responsibilities make it hard to maintain
+
+`CaffeineEntityCache` uses a single `Cache<EntityCacheRelationKey,
List<Entity>>` to store
+three semantically different kinds of data:
+
+| Stored data | Key form |
Example relation types |
+|-------------------------|--------------------------------------------------|-----------------------------------------------------------|
+| Direct entity | `(nameIdentifier, entityType, null)` |
any entity: catalog, schema, table, user, role, ... |
+| Relation result set | `(nameIdentifier, entityType, relType)` |
`ROLE_USER_REL`, `TAG_METADATA_OBJECT_REL`, ... |
+| Reverse index entries | `ReverseIndexCache` (separate radix tree) |
entity → list of cache keys that reference it |
+
+On top of this, a `cacheIndex` (radix tree) keeps a prefix-indexed view of all
keys to
+support cascading invalidation. The resulting invalidation logic
(`invalidateEntities`) is a
+BFS traversal that walks both the forward index and the reverse index, making
it difficult to
+reason about correctness and hard to extend safely.
+
+The five relation types currently tracked (`METADATA_OBJECT_ROLE_REL`,
`ROLE_USER_REL`,
+`ROLE_GROUP_REL`, `POLICY_METADATA_OBJECT_REL`, `TAG_METADATA_OBJECT_REL`) are
all
+auth-related, which reflects the original design intent: **the entity cache
was built
+primarily to serve the auth path.** Over time it accumulated relation types
and reverse-index
+logic without a clear ownership model, making it harder to maintain and evolve.
+
+##### Limited benefit for non-auth interfaces
+
+For general metadata API calls (list catalogs, list schemas, list tables), the
entity cache
+provides minimal benefit:
+
+| Operation | Goes through cache? | Notes
|
+|------------------------------------|---------------------|---------------------------------------------------|
+| `list(namespace, type)` | **No** | Bypasses cache
entirely; always hits DB |
+| `get(ident, type)` (single entity) | Yes | Cache helps on
repeated reads of the same entity |
+| `update(ident, type)` | Invalidate only | Invalidates
entry, write always goes to DB |
+| `listEntitiesByRelation(...)` | Yes | Only for the five
auth-centric relation types |
+
+In practice, the most common metadata browsing operations (`LIST` endpoints)
are not cached
+at the entity store level. The cache's real workload is the auth path, where
the same user
+entity, role assignments, and resource IDs are resolved on every single
authorization check.
+
+**Conclusion:** The entity cache is a de-facto auth cache dressed up as a
general-purpose
+cache. Its complexity is unjustified for the non-auth use case, and its
TTL-based consistency
+model is insufficient for the auth use case (see §1.8). A purpose-built auth
cache layer —
+separate from the entity store — is the cleaner path forward.
+
+---
+
+### 1.3 JCasbin Authorization — Deep Dive
+
+#### 1.3.1 Call Graph for a Single `authorize()` Check
+
+```
+JcasbinAuthorizer.authorize(principal, metalake, metadataObject, privilege)
+│
+├─ [A] getUserEntity(username, metalake)
+│ entityStore.get(USER by NameIdentifier)
+│ → Needed to obtain integer userId for JCasbin enforce()
+│
+├─ [B] MetadataIdConverter.getID(metadataObject, metalake) ← TARGET
RESOURCE
+│ entityStore.get(entity by NameIdentifier)
+│ → Needed to get integer metadataId for JCasbin enforce()
+│ → Called on every auth request
+│
+├─ [C] loadRolePrivilege(metalake, username, userId, requestContext)
+│ │ (guarded by requestContext.hasLoadRole — runs once per HTTP request)
+│ │
+│ ├─ [C1] entityStore.listEntitiesByRelation(ROLE_USER_REL, userIdentifier)
+│ │ → Get all roles assigned to this user
+│ │
+│ └─ For each role NOT already in loadedRoles cache:
+│ ├─ [C2] entityStore.get(RoleEntity by name) ← async, thread pool
+│ └─ loadPolicyByRoleEntity(roleEntity)
+│ └─ For each securableObject in role.securableObjects():
+│ ├─ [C3] MetadataIdConverter.getID(securableObject, metalake)
+│ └─ enforcer.addPolicy(roleId, objType, metadataId, privilege,
effect)
+│
+│ loadedRoles.put(roleId, true) ← mark role as loaded
+│
+├─ [D] isOwner() / loadOwnerPolicy(...) ← called on EVERY auth request (not
only OWNER
+│ │ privilege checks). Nearly all auth expressions contain ANY(OWNER,
METALAKE, CATALOG),
+│ │ which expands to METALAKE::OWNER || CATALOG::OWNER || … and calls
isOwner() directly
+│ │ via OGNL, independently of the authorize() path. Typical call count:
2–4 per request.
+│ ├─ Check ownerRel cache → if HIT, return (most non-owner users get
Optional.empty())
+│ └─ [D1] entityStore.listEntitiesByRelation(OWNER_REL, ...)
+│ ownerRel.put(metadataId, Optional.of(ownerId))
+│
+└─ [E] enforcer.enforce(userId, objectType, metadataId, privilege) ←
in-memory, O(1)
+```
+
+#### 1.3.2 What Each Cache Protects
+
+`loadedRoles: Cache<Long, Boolean>` — answers "is this role's policy already
in JCasbin?"
+Without it, every request re-executes [C2]+[C3] for all roles the user has
(N+1 queries).
+With it, [C2]+[C3] only run on first load per role. **This is the most
critical cache.**
+
+`ownerRel: Cache<Long, Optional<Long>>` — caches ownership lookups for
OWNER-privilege
+checks. **Contrary to initial analysis, `ownerRel` is consulted on virtually
every auth
+request**, not only when `privilege == OWNER`. The reason is that nearly every
authorization
+expression in `AuthorizationExpressionConstants` includes `ANY(OWNER,
METALAKE, CATALOG)`
+or similar clauses (e.g. `LOAD_TABLE_AUTHORIZATION_EXPRESSION`,
+`FILTER_TABLE_AUTHORIZATION_EXPRESSION`,
`LOAD_CATALOG_AUTHORIZATION_EXPRESSION`). The
+`ANY(OWNER, …)` macro expands to `METALAKE::OWNER || CATALOG::OWNER || …`, and
each
+`X::OWNER` term calls `isOwner()` directly — a code path that is **independent
of
+`authorize()`**. As a result, every auth request triggers 2–4 `isOwner()`
calls (one per
+ancestor level), each consulting `ownerRel`. For most non-owner users,
`ownerRel` caches
+`Optional.empty()`, which lets the ownership sub-check fail quickly without a
DB query.
+Without `ownerRel`, every auth request would add 2–4 extra DB queries against
`owner_meta`.
+
+**What these caches do NOT protect** (hit DB on every auth request without
entity cache):
+
+| Call | Description
| Protected by |
+|----------------------------------------------|-------------------------------------------|-------------------|
+| [A] `getUserEntity()` | Fetch User entity → get
integer userId | Entity cache only |
+| [B] `MetadataIdConverter.getID()` target | Resolve target resource name
→ integer ID | Entity cache only |
+| [C1] `listEntitiesByRelation(ROLE_USER_REL)` | Get user's role list
| Entity cache only |
+
+---
+
+### 1.4 Impact of Disabling Entity Cache
+
+Layer 2 sits **on top of** Layer 1. When Layer 1 is disabled (NoOpsCache),
calls [A], [B],
+[C1] hit DB on every auth request.
+
+| Call | With entity cache
| Without entity cache |
+|--------------------------------------------------|-------------------------------|---------------------------------|
+| [A] `getUserEntity()` | Cache hit after first
request | **DB query every auth request** |
+| [B] `MetadataIdConverter.getID()` target | Cache hit after first
request | **DB query every auth request** |
+| [C1] `listEntitiesByRelation(ROLE_USER_REL)` | Cache hit after first
request | **DB query every auth request** |
+| [C2] `entityStore.get(RoleEntity)` | Protected by
`loadedRoles` | DB only on cold role load |
+| [C3] `MetadataIdConverter.getID()` per privilege | Protected by
`loadedRoles` | DB only on cold role load |
+| [D1] `listEntitiesByRelation(OWNER_REL)` | Protected by `ownerRel`
| **DB query 2–4x per request** |
+
+---
+
+
+## 2. Goals
+
+### 2.1 The Two Problems to Solve
+
+**Problem 1 — Performance:** With entity cache disabled, [A] and [C1] hit DB
on every auth
+request. The new auth cache layer must protect these without relying on entity
store cache.
+([B] also hits DB, but this is correct and acceptable — see §1.5.)
+
+**Problem 2 — Consistency:** `loadedRoles` is TTL-bounded (1 hour staleness)
and updated by hook with in a instance. Permission
+changes must take effect at the next auth request, not after TTL expiry.
+
+Both problems are solved by the same mechanism: a version-validated cache for
the user's role
+list (userId comes for free from the same query).
+
+### 2.2 Requirements
+
+| Goal | Requirement
|
+|---------------------------------|---------------------------------------------------------------------------------------------------------------|
+| HA auth consistency | Privilege revocations visible on all nodes
at the next auth request |
+| Auth self-sufficiency | [A] and [C1] protected without relying on
entity store cache |
+| Auth performance | Hot path: ≤ 3 lightweight DB queries
|
+| No new mandatory infrastructure | Solution requires only the existing DB
|
+| Incremental delivery | Phase 1 independently shippable
|
+
+---
+
+## 3. Industry Reference
+
+### 3.1 Apache Polaris — Per-Entity Version Tracking
+
+#### Schema
+
+All entity types (catalogs, namespaces, tables, roles, principals) share a
single `ENTITIES`
+table (single-table inheritance). The two version columns are the key fields
for caching:
+
+```sql
+ENTITIES (
+ id BIGINT, -- Unique entity ID
+ catalog_id BIGINT, -- Owning catalog (0 for top-level entities)
+ parent_id BIGINT, -- Parent entity ID, forms the hierarchy
tree
+ type_code INT, -- Entity type enum (see hierarchy below)
+ name VARCHAR,
+ entity_version INT, -- Bumped on rename / property update /
drop ← key
+ sub_type_code INT, -- Subtype (ICEBERG_TABLE, ICEBERG_VIEW,
etc.)
+ properties JSON, -- User-visible properties (location,
format, etc.)
+ internal_properties JSON, -- Internal properties (credentials,
storage config, etc.)
+ grant_records_version INT, -- Bumped on every GRANT or REVOKE
← key
+)
+
+GRANT_RECORDS (
+ securable_catalog_id BIGINT,
+ securable_id BIGINT, -- The resource being secured
(table/namespace/catalog)
+ grantee_catalog_id BIGINT,
+ grantee_id BIGINT, -- The principal or role receiving the grant
+ privilege_code INT -- One of 102 defined privileges
+)
+```
+
+`GRANT_RECORDS` has no version column of its own. The version fingerprint is
stored in
+`ENTITIES.grant_records_version` — detecting staleness requires no scan of
`GRANT_RECORDS`.
+
+#### Entity Type Hierarchy
+
+```
+ROOT
+ ├── PRINCIPAL (user account, isGrantee)
+ ├── PRINCIPAL_ROLE (user-level role, isGrantee)
+ └── CATALOG
+ ├── CATALOG_ROLE (catalog-level role, isGrantee)
+ ├── NAMESPACE
+ │ └── TABLE_LIKE / POLICY / FILE
+ └── TASK
+```
+
+Only `PRINCIPAL`, `PRINCIPAL_ROLE`, and `CATALOG_ROLE` are **grantees** (can
receive grants).
+All others are **securables** (privileges are set on them).
+
+#### How `grantRecordsVersion` Is Maintained
+
+Every `grantPrivilege` / `revokePrivilege` call performs three writes in **one
DB transaction**:
+
+1. Insert or delete the `GRANT_RECORDS` row.
+2. Increment `grant_records_version` on the **grantee** entity row.
+3. Increment `grant_records_version` on the **securable** entity row.
+
+Both sides are bumped atomically — no separate changelog table is needed.
+
+#### Version-Validated Cache
+
+The cache unit is `ResolvedPolarisEntity` = entity metadata + grant records in
both directions.
+On every request, `bulkValidate()` issues one batch query for all path
entities:
+
+```sql
+SELECT * FROM ENTITIES WHERE (catalog_id, id) IN ((?, ?), ...)
+```
+
+| Path | Condition | Action
|
+|-------------------------|------------------------|----------------------------------------|
+| Cache hit | Both versions current | Serve from cache — **0
extra queries** |
+| Stale, targeted refresh | Either version behind | Reload only the changed
dimension |
+| Cache miss | Not in cache | Full load
|
+
+The DB is the single source of truth; no broadcast is needed for correctness.
+
+**Key difference from Gravitino:** Polaris bundles entity + grants in one
cached object, so one
+batch query covers both dimensions. Gravitino separates user→role from
role→privilege, requiring
+2 version-check queries on a warm hit (see §4.1 Step 1 and Step 3). Both
achieve strong
+consistency.
+
+### 3.2 Other References
+
+**Nessie** — HTTP fan-out invalidation: async POST to peer nodes on write,
convergence < 200 ms.
+
+**Keycloak** — JGroups embedded cluster messaging: in-JVM broadcast, no
separate service.
+Recommended future direction if Gravitino needs stronger delivery guarantees.
+
+**DB version polling** — monotonic counters incremented in write transaction;
a background
+thread polls for version changes and proactively invalidates caches.
Considered but not
+adopted; per-request validation (§4.1) achieves strong consistency without
background threads.
+
+---
+
+## 4. Design
+
+### 4.1 Per-Request Version Check (Polaris Style)
+
+Every auth request executes two lightweight version-check queries before
serving from cache.
+If any version has advanced, only the stale portion is reloaded. Staleness
window: **zero**.
+
+#### 4.1.1 Schema Changes
+
+Three new version columns, all `DEFAULT 1` — fully backward compatible.
Existing rows get
+version 1; first auth check after migration populates caches normally.
+
+```sql
+ALTER TABLE `role_meta`
+ ADD COLUMN `securable_objects_version` INT UNSIGNED NOT NULL DEFAULT 1
+ COMMENT 'Incremented atomically with any privilege grant/revoke for this
role';
+
+ALTER TABLE `user_meta`
+ ADD COLUMN `role_grants_version` INT UNSIGNED NOT NULL DEFAULT 1
+ COMMENT 'Incremented atomically with any role assignment/revocation for
this user';
+
+ALTER TABLE `group_meta`
+ ADD COLUMN `role_grants_version` INT UNSIGNED NOT NULL DEFAULT 1
+ COMMENT 'Incremented atomically with any role assignment/revocation for
this group';
+```
+
+Write paths that must bump the version **in the same DB transaction**:
+
+| Operation | Column
| Location |
+|------------------------------------|---------------------------------------------------------|--------------------|
+| Grant / revoke privilege on role R | `role_meta.securable_objects_version
WHERE role_id = R` | `RoleMetaService` |
+| Assign / revoke role for user U | `user_meta.role_grants_version WHERE
user_id = U` | `UserMetaService` |
+| Assign / revoke role for group G | `group_meta.role_grants_version WHERE
group_id = G` | `GroupMetaService` |
+
+Version comparison uses `!=` (not `<`) to safely handle theoretical INT
wrap-around.
+
+**Ownership transfers** require no schema change and no cache. The `ownerRel`
cache is
+**removed** (see §7.2). Step 2.5 queries `owner_meta` directly with a single
batch query on
+every auth request — strongly consistent, no versioning complexity needed.
+
+#### 4.1.2 Cache Data Structures (Changes in JcasbinAuthorizer)
+
+```java
+// ─── BEFORE ──────────────────────────────────────────────────────────
+private Cache<Long, Boolean> loadedRoles; // roleId → loaded?
+private Cache<Long, Optional<Long>> ownerRel;
+
+// ─── AFTER ───────────────────────────────────────────────────────────
+
+// NEW: replaces entity cache dependency for [A] (userId) and [C1] (role list).
+// Step 1 query returns both user_id and role_grants_version in one shot.
+// metalakeName→metalakeId resolved inline via JOIN — no dedicated cache
needed.
+private GravitinoCache<String, CachedUserRoles> userRoleCache;
+// key = metalakeName + ":" + userName
+
+record CachedUserRoles(
+ long userId, // integer userId for JCasbin enforce()
+ int roleGrantsVersion, // user_meta.role_grants_version at load time
+ List<Long> roleIds // role ID list at load time
+) {}
+
+// NEW: mirrors userRoleCache for groups (group can also hold role
assignments).
+private GravitinoCache<String, CachedGroupRoles> groupRoleCache;
+// key = metalakeName + ":" + groupName
+
+record CachedGroupRoles(
+ long groupId,
+ int roleGrantsVersion, // group_meta.role_grants_version at load
time
+ List<Long> roleIds
+) {}
+
+// TYPE CHANGE: was Cache<Long, Boolean>, now stores securable_objects_version.
+// Enables version-based staleness detection rather than TTL expiry.
+private GravitinoCache<Long, Integer> loadedRoles;
+// roleId → securable_objects_version at the time JCasbin policies were loaded
+
+// REMOVED: ownerRel cache eliminated (see §7.2).
+// isOwner() is called 2–4 times per request, but version-validated caching
offers no
+// query savings: the version check query already returns the owner_id, so
there is nothing
+// expensive to avoid. Step 2.5 queries owner_meta directly via one batch
query per request.
+// private Cache<Long, Optional<Long>> ownerRel;
+```
+
+**Why no cache for [B] (target resource name→ID):**
+Adding a `metadataIdCache` would require invalidation on every entity rename,
drop, or
+recreate across all entity types. Since JCasbin uses integer IDs (not names),
the DB lookup
+for [B] is always correct (~1 ms indexed). Simpler and more correct to hit DB
every request.
+
+**Why `ownerRel` is removed:**
+`isOwner()` is called 2–4 times per auth request (once per ancestor in the
OGNL expression
+chain). A version-validated cache would still query `owner_meta` on every
request to check
+versions — and that same query already returns the `owner_id`. There is no
expensive
+downstream work to avoid (unlike roles, where skipping securable-object
loading saves
+significant work). Caching adds complexity with zero query savings. Step 2.5
issues one
+batch query per request directly against `owner_meta`, achieving strong
consistency with
+no extra infrastructure. See §7.2.
Review Comment:
Most of the DB operations require joining multiple tables, I'm not sure your
assumption of DB access (1ms) is valid or not.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
