[HWORKS-2802 / -2807] Document partitioned_by on feature group creation

docs/user_guides/fs/feature_group/create.md

@@ -102,6 +102,59 @@ MaxDirectoryItemsExceededException - The directory item limit is exceeded: limit
 
 By using partitioning the system will write the feature data in different subdirectories, thus allowing you to write 10240 files per partition.
 
+##### Time-grain partitioning with `partitioned_by` (Delta only)
+
+When the partition columns are derived from the feature group's `event_time`, hand the backend the desired time grains with `partitioned_by=[...]` and the Python client derives the partition columns for you.
+Pass one or more grains drawn from `hour`, `day`, `week`, `month`, and `year`.
+
+```python
+fg = fs.get_or_create_feature_group(
+    name="transactions",
+    version=1,
+    primary_key=["tx_id"],
+    event_time="tx_ts",
+    partitioned_by=["year", "month", "day"],
+    time_travel_format="DELTA",
+)
+fg.insert(df)  # df does not need year/month/day — the client derives them
+```
+
+The example above is equivalent to manually decomposing `tx_ts` into three columns and passing `partition_key=["year", "month", "day"]`.
+The grain columns are ordinary materialized partition columns: the client computes them from `event_time` on each write and the backend registers them as partition columns through the normal table-creation path.
+The source dataframe does not need to carry them.
+
+`partitioned_by` and `partition_key` are mutually exclusive.
+`partitioned_by` requires `event_time` to be set.
+
+###### Partition pruning
+
+The grain columns are real partition columns, so a filter on a grain column (for example `year == 2026`) prunes partitions natively.
+A filter on an `event_time` range is rewritten into equivalent grain-column predicates by the query layer, so it prunes too on hierarchical specs:
+
+| `partitioned_by` | Prunes on `event_time` range? | Prunes on `year` / `month` / `day` filter? |
+| --- | --- | --- |
+| `["year"]` | ✅ | ✅ |
+| `["year", "month"]` | ✅ | ✅ |
+| `["year", "month", "day"]` | ✅ | ✅ |
+| `["year", "month", "day", "hour"]` | ✅ | ✅ |
+| `["month"]` (no year) | ⚠️ no — month alone is ambiguous across years | ✅ filter on month works |
+| `["year", "week"]` | ⚠️ year only — week isn't directly derivable from a date range | ✅ both columns prune |
+| `["day"]` (no year/month) | ⚠️ no — day-of-month is ambiguous | ✅ filter on day works |
+
+Prefer hierarchical specs (`["year"]`, `["year", "month"]`, `["year", "month", "day"]`) — they line up with the typical batch-pipeline access pattern and prune naturally.
+
+###### Online feature store
+
+Online-enabled feature groups do not yet support `partitioned_by`.
+The online ingestion path does not exclude the offline-only grain columns from the Kafka/Avro schema, nor materialize them for the online write, so the backend rejects `partitioned_by` together with `online_enabled=true` until that work lands (tracked under a separate follow-up ticket).
+Keep the feature group offline-only to use `partitioned_by`.
+
+###### Hudi
+
+`partitioned_by` on `time_travel_format="HUDI"` feature groups is not yet supported and the backend rejects it at creation.
+Hudi materializes the grain columns server-side in the streaming materialization job, and that work is tracked under a separate follow-up ticket.
+Until that lands, use `time_travel_format="DELTA"` to get time-grain partitioning, or partition Hudi groups explicitly via `partition_key=["year"]` with a `year` column the upstream pipeline computes.
+
 ##### Table format
 
 When you create a feature group, you can specify the table format you want to use to store the data in your feature group by setting the `time_travel_format` parameter.