twalthr commented on code in PR #29090:
URL: https://github.com/apache/flink/pull/29090#discussion_r3932707744


##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.
+
+## The four row kinds
+
+Every row in a changelog stream has a `RowKind`:
+
+| Kind | Symbol | Meaning |
+|---|---|---|
+| `INSERT` | `+I` | A new row was added. |
+| `UPDATE_BEFORE` | `-U` | The previous value of a row that is about to be 
updated. |
+| `UPDATE_AFTER` | `+U` | The new value of a row that was updated. |
+| `DELETE` | `-D` | A row was removed. |
+
+## Changelog Modes
+
+Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on the operations 
present in your query and each of these modes has it's own characteristic.

Review Comment:
   ```suggestion
   Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on source 
connector and the operations present in your query. Each of these modes has 
it's own characteristic:
   ```



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.
+
+## The four row kinds
+
+Every row in a changelog stream has a `RowKind`:
+
+| Kind | Symbol | Meaning |
+|---|---|---|
+| `INSERT` | `+I` | A new row was added. |
+| `UPDATE_BEFORE` | `-U` | The previous value of a row that is about to be 
updated. |
+| `UPDATE_AFTER` | `+U` | The new value of a row that was updated. |
+| `DELETE` | `-D` | A row was removed. |
+
+## Changelog Modes
+
+Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on the operations 
present in your query and each of these modes has it's own characteristic.
+
+**Append Mode `{+I}`**
+- All messages are insert-only.
+- Every insertion message is an immutable fact.
+- Messages can be distributed in an arbitrary fashion across partitions and 
processors because they are unrelated.
+
+**Upsert Mode `{+I, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Updates are related using a key (i.e. the *upsert key*).
+- Every message is either an upsert or delete message for a result under the 
upsert key.
+- Messages for the same upsert key should land at the same partition and 
processor.
+- Deletions can contain only values for upsert key columns (i.e. *partial 
deletes*) or values for
+  all columns (i.e. *full deletes*).
+- The mode is also known as *partial image* in the literature because `-U` 
messages are missing.
+
+**Retract Mode `{+I, -U, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Every insertion or update event is a fact that can be "undone" (i.e. 
retracted).
+- Updates are related by all columns. In simplified words: The entire row is 
kind of the key but duplicates are supported.
+  For example: `+I['Bob', 42]` is related to `-D['Bob', 42]` and `+U['Alice', 
13]` is related to `-U['Alice', 13]`.
+- Thus, every message is either an insertion (`+`) or its retraction (`-`).
+- The mode is known as *full image* in the literature.
+
+## Reading it as a consumer
+
+To read a changelog correctly, you don't need to figure out which mode you're 
in — check one piece of metadata: whether the table has a primary key.
+
+**If you have a primary key**, materialize rows in a map keyed by it:
+
+- `+I`: add the row for that key.
+- `+U`: replace the row for that key (add it, if you haven't seen that key 
before).
+- `-D`: remove the row for that key.
+- `-U`: ignore it. The `+U` for the same key replaces it right after.
+
+**If you don't have a primary key**, there's no key to look up — match on the 
full row instead:
+
+- `+I` or `+U`: add the row.
+- `-U` or `-D`: remove the row with exactly this content. Flink sends the 
exact previous values on `-U`, so it matches the row you already added.
+
+## Valid Changelog Examples
+
+### With a primary key
+
+#### Insert

Review Comment:
   Maybe add a bit explanation to each example?



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.
+
+## The four row kinds
+
+Every row in a changelog stream has a `RowKind`:
+
+| Kind | Symbol | Meaning |
+|---|---|---|
+| `INSERT` | `+I` | A new row was added. |
+| `UPDATE_BEFORE` | `-U` | The previous value of a row that is about to be 
updated. |
+| `UPDATE_AFTER` | `+U` | The new value of a row that was updated. |
+| `DELETE` | `-D` | A row was removed. |
+
+## Changelog Modes
+
+Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on the operations 
present in your query and each of these modes has it's own characteristic.
+
+**Append Mode `{+I}`**
+- All messages are insert-only.
+- Every insertion message is an immutable fact.
+- Messages can be distributed in an arbitrary fashion across partitions and 
processors because they are unrelated.
+
+**Upsert Mode `{+I, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Updates are related using a key (i.e. the *upsert key*).
+- Every message is either an upsert or delete message for a result under the 
upsert key.
+- Messages for the same upsert key should land at the same partition and 
processor.
+- Deletions can contain only values for upsert key columns (i.e. *partial 
deletes*) or values for
+  all columns (i.e. *full deletes*).
+- The mode is also known as *partial image* in the literature because `-U` 
messages are missing.
+
+**Retract Mode `{+I, -U, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Every insertion or update event is a fact that can be "undone" (i.e. 
retracted).
+- Updates are related by all columns. In simplified words: The entire row is 
kind of the key but duplicates are supported.
+  For example: `+I['Bob', 42]` is related to `-D['Bob', 42]` and `+U['Alice', 
13]` is related to `-U['Alice', 13]`.
+- Thus, every message is either an insertion (`+`) or its retraction (`-`).
+- The mode is known as *full image* in the literature.
+
+## Reading it as a consumer
+
+To read a changelog correctly, you don't need to figure out which mode you're 
in — check one piece of metadata: whether the table has a primary key.
+
+**If you have a primary key**, materialize rows in a map keyed by it:
+
+- `+I`: add the row for that key.
+- `+U`: replace the row for that key (add it, if you haven't seen that key 
before).
+- `-D`: remove the row for that key.
+- `-U`: ignore it. The `+U` for the same key replaces it right after.
+
+**If you don't have a primary key**, there's no key to look up — match on the 
full row instead:
+
+- `+I` or `+U`: add the row.
+- `-U` or `-D`: remove the row with exactly this content. Flink sends the 
exact previous values on `-U`, so it matches the row you already added.
+
+## Valid Changelog Examples
+
+### With a primary key
+
+#### Insert
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Update, upsert shape
+
+```
+Flink Changelog
+-- +U[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Update, retract shape
+
+```
+Flink Changelog
+-- -U[ id: 5, name: 'old' ]
+-- +U[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Delete
+
+```
+Flink Changelog
+-- -D[ id: 5, name: 'new' ]
+
+Table
+(empty)
+```
+
+#### Update with partial delete
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'a' ]
+-- +U[ id: 5, name: 'b' ]
+-- -D[ id: 5, name: null ]
+
+Table
+(empty)
+```
+
+The `-D` only carries `id`; `name` arrives as `null` rather than `'b'`. The 
row is still removed correctly because removal is driven by the key alone. The 
engine tries to do this optimization when possible. As a consumer you should be 
aware: if you have a key, you might receive deletions with a complete or 
partial delete.
+
+#### Distributed across partitions
+
+Flink partitions the changelog by the primary key, so every event for a given 
key always lands on the same partition, in order:
+
+```
+Flink Changelog
+
+Partition 0
+-- -U[ id: 5, name: 'old' ]
+-- +U[ id: 5, name: 'new' ]
+
+Partition 1
+-- +I[ id: 9, name: 'x' ]
+-- -D[ id: 9, name: 'x' ]
+
+Table, read while Partition 1's +I has landed but not yet its -D (transient)
+-- [ id: 5, name: 'new' ]
+-- [ id: 9, name: 'x' ]
+
+Table, read after both partitions have delivered (correct)
+-- [ id: 5, name: 'new' ]
+```
+
+Order is guaranteed within a partition, not across partitions. `id: 5`'s 
update and `id: 9`'s insert-then-delete never have to be reconciled with each 
other — each key's full history stays on one partition, in order. But 
partitions are still processed independently of each other, so reading the 
table mid-way, between Partition 1's two events, briefly shows `id: 9` even 
though it's ultimately deleted. That's transient, not a bug — it resolves once 
Partition 1's `-D` lands.
+
+### Without a primary key
+
+#### Insert
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Update
+
+```
+Flink Changelog
+-- +U[ id: 5, name: 'old' ]
+-- -U[ id: 5, name: 'old' ]
+-- +U[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Delete
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'new' ]
+-- -D[ id: 5, name: 'new' ]
+
+Table
+(empty)
+```
+
+#### Distributed across partitions
+
+Without a key, Flink partitions by the entire row instead — so only an event 
and an exact match of it (its own retraction, its own repeat) are guaranteed to 
land on the same partition. Two rows that happen to share `id` but differ 
elsewhere can land anywhere:

Review Comment:
   This is not fully correct. We still partition by the GROUP BY or PARTITION 
BY key.



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.

Review Comment:
   (needs a bit of polishing and cross reference links)



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.
+
+## The four row kinds
+
+Every row in a changelog stream has a `RowKind`:
+
+| Kind | Symbol | Meaning |
+|---|---|---|
+| `INSERT` | `+I` | A new row was added. |
+| `UPDATE_BEFORE` | `-U` | The previous value of a row that is about to be 
updated. |
+| `UPDATE_AFTER` | `+U` | The new value of a row that was updated. |
+| `DELETE` | `-D` | A row was removed. |
+
+## Changelog Modes
+
+Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on the operations 
present in your query and each of these modes has it's own characteristic.
+
+**Append Mode `{+I}`**
+- All messages are insert-only.
+- Every insertion message is an immutable fact.
+- Messages can be distributed in an arbitrary fashion across partitions and 
processors because they are unrelated.
+
+**Upsert Mode `{+I, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Updates are related using a key (i.e. the *upsert key*).
+- Every message is either an upsert or delete message for a result under the 
upsert key.
+- Messages for the same upsert key should land at the same partition and 
processor.
+- Deletions can contain only values for upsert key columns (i.e. *partial 
deletes*) or values for
+  all columns (i.e. *full deletes*).
+- The mode is also known as *partial image* in the literature because `-U` 
messages are missing.
+
+**Retract Mode `{+I, -U, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Every insertion or update event is a fact that can be "undone" (i.e. 
retracted).
+- Updates are related by all columns. In simplified words: The entire row is 
kind of the key but duplicates are supported.
+  For example: `+I['Bob', 42]` is related to `-D['Bob', 42]` and `+U['Alice', 
13]` is related to `-U['Alice', 13]`.
+- Thus, every message is either an insertion (`+`) or its retraction (`-`).
+- The mode is known as *full image* in the literature.
+
+## Reading it as a consumer
+
+To read a changelog correctly, you don't need to figure out which mode you're 
in — check one piece of metadata: whether the table has a primary key.
+
+**If you have a primary key**, materialize rows in a map keyed by it:
+
+- `+I`: add the row for that key.
+- `+U`: replace the row for that key (add it, if you haven't seen that key 
before).
+- `-D`: remove the row for that key.
+- `-U`: ignore it. The `+U` for the same key replaces it right after.
+
+**If you don't have a primary key**, there's no key to look up — match on the 
full row instead:
+
+- `+I` or `+U`: add the row.
+- `-U` or `-D`: remove the row with exactly this content. Flink sends the 
exact previous values on `-U`, so it matches the row you already added.
+
+## Valid Changelog Examples
+
+### With a primary key
+
+#### Insert
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Update, upsert shape
+
+```
+Flink Changelog
+-- +U[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Update, retract shape
+
+```
+Flink Changelog
+-- -U[ id: 5, name: 'old' ]
+-- +U[ id: 5, name: 'new' ]
+
+Table
+-- [ id: 5, name: 'new' ]
+```
+
+#### Delete
+
+```
+Flink Changelog
+-- -D[ id: 5, name: 'new' ]
+
+Table
+(empty)
+```
+
+#### Update with partial delete
+
+```
+Flink Changelog
+-- +I[ id: 5, name: 'a' ]
+-- +U[ id: 5, name: 'b' ]
+-- -D[ id: 5, name: null ]
+
+Table
+(empty)
+```
+
+The `-D` only carries `id`; `name` arrives as `null` rather than `'b'`. The 
row is still removed correctly because removal is driven by the key alone. The 
engine tries to do this optimization when possible. As a consumer you should be 
aware: if you have a key, you might receive deletions with a complete or 
partial delete.
+
+#### Distributed across partitions
+
+Flink partitions the changelog by the primary key, so every event for a given 
key always lands on the same partition, in order:

Review Comment:
   ```suggestion
   Flink partitions the changelog by the upsert key, so every event for a given 
key always lands on the same partition, in order:
   ```



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.

Review Comment:
   ```suggestion
   A Flink table can change over time: rows get inserted, updated, or deleted. 
Internally, the Flink engine encodes this as streams containing changelogs. 
Those changelogs are typically hidden from a SQL user.
   
   However, advanced API endpoints reveal the inner workings:
   - When implementing custom connectors
   - A PTF that consumes or produces changelogs
   - When using TO_CHANGELOG or FROM_CHANGELOG functions
   - When brinding to DataStream API (toChangelogStream, fromChangelogStream)
   
   
   This page describes how to interpret that stream correctly for advanced use 
cases.
   ```



##########
docs/content/docs/concepts/sql-table-concepts/interpreting_changelog.md:
##########
@@ -0,0 +1,326 @@
+---
+title: "Interpreting Flink's Changelog"
+weight: 3
+type: docs
+---
+<!--
+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.
+-->
+
+# Interpreting Flink's Changelog
+
+A Flink table can change over time: rows get inserted, updated, or deleted. 
This is internally encoded internally as changelogs inside the Flink SQL 
engine. When you consume that table as a stream — after writing the table to 
another system, through `toChangelogStream`, a table connector, or your own 
service reading the output — each row carries a change flag telling you what 
kind of change it represents. This page describes how to read that stream 
correctly so you're able to built a table representation out of it.
+
+## The four row kinds
+
+Every row in a changelog stream has a `RowKind`:
+
+| Kind | Symbol | Meaning |
+|---|---|---|
+| `INSERT` | `+I` | A new row was added. |
+| `UPDATE_BEFORE` | `-U` | The previous value of a row that is about to be 
updated. |
+| `UPDATE_AFTER` | `+U` | The new value of a row that was updated. |
+| `DELETE` | `-D` | A row was removed. |
+
+## Changelog Modes
+
+Under the hood, the Flink SQL engine is always operating with changelogs in 
one of three modes: append, upsert and retract. This depends on the operations 
present in your query and each of these modes has it's own characteristic.
+
+**Append Mode `{+I}`**
+- All messages are insert-only.
+- Every insertion message is an immutable fact.
+- Messages can be distributed in an arbitrary fashion across partitions and 
processors because they are unrelated.
+
+**Upsert Mode `{+I, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Updates are related using a key (i.e. the *upsert key*).
+- Every message is either an upsert or delete message for a result under the 
upsert key.
+- Messages for the same upsert key should land at the same partition and 
processor.
+- Deletions can contain only values for upsert key columns (i.e. *partial 
deletes*) or values for
+  all columns (i.e. *full deletes*).
+- The mode is also known as *partial image* in the literature because `-U` 
messages are missing.
+
+**Retract Mode `{+I, -U, +U, -D}`**
+- Messages can contain updates leading to an updating table.
+- Every insertion or update event is a fact that can be "undone" (i.e. 
retracted).
+- Updates are related by all columns. In simplified words: The entire row is 
kind of the key but duplicates are supported.
+  For example: `+I['Bob', 42]` is related to `-D['Bob', 42]` and `+U['Alice', 
13]` is related to `-U['Alice', 13]`.
+- Thus, every message is either an insertion (`+`) or its retraction (`-`).
+- The mode is known as *full image* in the literature.
+
+## Reading it as a consumer
+
+To read a changelog correctly, you don't need to figure out which mode you're 
in — check one piece of metadata: whether the table has a primary key.

Review Comment:
   ```suggestion
   To read a changelog correctly, you don't need to figure out which mode 
you're in — check one piece of metadata: whether the table has a upsert key.
   ```
   maybe introduce the concept of an upsert key? At the source and sink it 
matches the primary key, but in queries are usually the result of a join key, 
GROUP BY, or PARTITION BY.



-- 
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