jerryshao commented on code in PR #12157:
URL: https://github.com/apache/gravitino/pull/12157#discussion_r3664382678


##########
design-docs/treelock-necessity-and-concurrency-design.md:
##########
@@ -0,0 +1,528 @@
+<!--
+  Licensed to the Apache Software Foundation (ASF) under one
+  or more contributor license agreements. See the NOTICE file
+  distributed with this work for additional information
+  regarding copyright ownership. The ASF licenses this file
+  to you under the Apache License, Version 2.0 (the
+  "License"); you may not use this file except in compliance
+  with the License. You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+  Unless required by applicable law or agreed to in writing,
+  software distributed under the License is distributed on an
+  "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+  KIND, either express or implied. See the License for the
+  specific language governing permissions and limitations
+  under the License.
+-->
+
+# Design: Concurrency Control for Multi-Node Gravitino (TreeLock)
+
+> Tracking issue: [#10474](https://github.com/apache/gravitino/issues/10474) — 
*Address TreeLock limitations for Gravitino HA deployment*
+
+**Short words used in this doc:**
+
+- **HA** = High Availability = running more than one Gravitino server at the 
same time behind a load balancer.
+- **OCC** = Optimistic Concurrency Control = the caller does not take an 
application/path lock before work; when it writes, it checks "is the row still 
the version I read?" If yes, write; if no, someone else changed it first, so 
handle the conflict. The database still takes its normal short 
statement/transaction locks.
+- **External catalog** = the real system that holds the data, such as Hive, 
Iceberg, MySQL, or Kafka.
+- **Gravitino store** = Gravitino's own database (`RelationalEntityStore`) 
that keeps a copy of the metadata.
+- **Source of truth** = the system whose data we trust as correct when two 
copies disagree.
+
+---
+
+## Background
+
+Gravitino uses an in-memory lock called `TreeLock` 
(`core/src/main/java/org/apache/gravitino/lock/`) to run metadata operations 
one at a time. For every operation, `TreeLockUtils.doWithTreeLock` locks the 
whole path from the root down: a read lock on every parent, and a read or write 
lock on the target (or on its parent for rename/drop).
+
+```
+loadTable(metalake.cat.db.t1)        alterTable(... rename t1)        
dropTable(metalake.cat.db.t1)
+  /                  READ              /                  READ          /      
            READ
+  /metalake          READ             /metalake          READ          
/metalake          READ
+  /metalake/cat      READ             /metalake/cat      READ          
/metalake/cat      READ
+  /metalake/cat/db   READ             /metalake/cat/db   WRITE         
/metalake/cat/db   WRITE
+  /metalake/cat/db/t1 READ            (parent write-locked)            (parent 
write-locked)
+```
+
+This is built on `LockManager`, which keeps an in-memory tree of 
`TreeLockNode`s (each one wraps a `ReentrantReadWriteLock`), plus reference 
counting, a background thread that removes unused nodes, and another background 
thread that checks for deadlocks inside the same JVM. It is about 800 lines of 
code in total (`LockManager` ~300, `TreeLockNode` ~250, `TreeLock` ~180, 
`TreeLockUtils` ~70).
+
+### The problem: TreeLock only works inside one JVM
+
+Each Gravitino server has its **own** `LockManager` with its own lock tree in 
its own memory. A write lock taken on server A means nothing to server B. 
Behind a load balancer, two servers can each pass their *local* TreeLock and 
change the same resource at the same time. So the moment Gravitino runs in HA, 
TreeLock stops protecting anything across servers. This is the reason #10474 
was opened.
+
+This leaves us with a decision. There are two directions:
+
+1. **Remove TreeLock's correctness role** and let explicit rules in the shared 
database keep data correct.
+2. **Keep the lock but make it work across nodes** (a distributed lock).
+
+The rest of this document analyses what TreeLock really does today, then 
compares these two directions, then picks one.
+
+---
+
+## Goals
+
+1. **Understand what TreeLock protects today**: Describe what TreeLock 
actually guards, and which of those guarantees the shared database could 
provide instead.
+2. **Evaluate the candidate directions on equal footing**: Assess each 
direction — database-native concurrency, and a cross-node distributed lock — 
against the same criteria (correctness, performance, maintainability, 
operational cost), informed by how comparable systems solve the two-store 
problem. Neither direction is assumed better going in.
+3. **Correctness for one and for many servers**: Whatever is chosen must be 
correct with a single server and under HA.
+4. **Decide with evidence**: End with a direction and the reasons behind it, 
plus a plan a developer can start on.
+
+---
+
+## Non-Goals
+
+1. **No single transaction across the external catalog and the Gravitino 
store**: We will not try 2PC/XA across the two stores. External catalogs do not 
all offer the same transaction guarantees, so the existing pattern (re-sync on 
read, which is safe to repeat) is kept.
+2. **Write correctness only, not read staleness**: This document is about 
correct concurrent *writes*. Keeping each server's *cached reads* fresh across 
nodes (the `EntityChangeLogPoller` work) is a separate effort, out of scope 
here.
+3. **No change to external-catalog behavior**: We will not change how 
Hive/Iceberg/JDBC connectors keep their own data correct; they are the source 
of truth for their own data.
+
+---
+
+## Analysis and Investigation
+
+### A metadata write touches two stores, with no shared transaction
+
+The key fact that is easy to miss: **every catalog metadata operation touches 
two separate stores, and there is no single transaction that covers both of 
them**:
+
+1. The **external catalog** (Hive, Iceberg REST, JDBC/MySQL, Kafka, …).
+2. Gravitino's **own database**, the Gravitino store.
+
+The order is always the same: **the external system first, the Gravitino store 
second.** From `TableOperationDispatcher`:
+
+```text
+internalCreateTable():  catalog.createTable(...)   →  store.put(tableEntity)   
  // lines 642-689
+dropTable():            catalog.dropTable(...)      →  store.delete(ident, 
TABLE) // lines 366-386
+alterTable():           catalog.alterTable(...)     →  store column-sync       
   // lines 267-340
+importTable():          catalog.loadTable(...)      →  store.put(tableEntity)  
   // lines 474-527
+```
+
+![Two-store write and the crash gap](images/treelock-two-store-write.png)
+
+When a table is created, its Gravitino id is written **into the external 
table's own properties** as a `StringIdentifier` (`internalCreateTable`, line 
636). Later, when the table is read, `importTable` uses the data from the 
external system to **overwrite and correct** the stored copy. So for 
external-backed catalogs, the external system is the source of truth, and the 
Gravitino store is a copy that is updated later and fixes itself on the next 
read.
+
+An important result: **no lock can make the two stores update as one unit.** 
If the process crashes after `catalog.createTable` finished but before 
`store.put` runs, the external system has a table with no matching Gravitino 
entity. This is a crash problem, not a "two things at once" problem, and 
neither a local nor a distributed lock can fix it. This already tells us that a 
lock is not the tool that keeps the two stores matched.
+
+### How the store fixes itself, and where it stops working
+
+Here is how the self-fix works, traced through the code. If create succeeds in 
the external system but the store write fails, the error is **hidden** — 
`internalCreateTable:688-696` catches it, logs it, and still returns success to 
the client with no stored entity. Nothing is fixed until someone reads the 
table again. On that next read, `loadTable:142` sees `imported == false` and 
runs `importTable`, which writes the entity into the store. Schemas behave the 
same way (`internalLoadSchema` + `importSchema`). So a failed store write 
**does** auto-correct on the next load.
+
+![How import decides between reusing the id and assigning a new 
id](images/treelock-import-decision.png)
+
+The import step does not need the id to run; the id only decides how stable 
the result is. The details:
+
+| External system                                                         | 
How "needs import" is detected                                                 
| Id after the self-fix                                                       |
+| ----------------------------------------------------------------------- | 
------------------------------------------------------------------------------ 
| --------------------------------------------------------------------------- |
+| Can store the id (e.g. Hive, Iceberg) — `stringId != null`              | by 
id — `store.get(stringId.id())` (`internalLoadTable:596`)                   | 
**the original id is reused** (`importTable` sets `uid = stringId.id()`)    |
+| Cannot store the id (e.g. JDBC schema, PostgreSQL) — `stringId == null` | by 
name — `getEntity(ident)` (`internalLoadTable:570`, code comment at `:590`) | 
**a new id is generated** (`importTable` sets `uid = idGenerator.nextId()`) |
+
+The narrow limit: for an id-less external system the id is really owned by the 
Gravitino store, so if that row is lost the id cannot be recovered, and 
anything that references the entity by id (owner, tag, policy, role) points at 
the old, now-missing id. This does not appear in the plain "create then store 
write fails" flow, because no id-based references exist yet.
+
+### Not every catalog has an external system that decides the winner

Review Comment:
   For the self managed entities, we should guarantee the strong consistency.



-- 
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]

Reply via email to