rdblue commented on code in PR #16425:
URL: https://github.com/apache/iceberg/pull/16425#discussion_r4649419183
##########
format/spec.md:
##########
@@ -1801,6 +1805,81 @@ If a delete column in an equality delete file is later
dropped from the table, i
Manifests hold the same statistics for delete files and data files. For delete
files, the metrics describe the values that were deleted.
+### Column Files
+
+A data file or leaf manifest may store fields in column files listed in
[`column_files`](#column-file-fields). The data file or leaf manifest whose
`column_files` lists a column file is that column file's _base file_. A column
file can use any supported Iceberg file format and must have the same row count
as its base file.
+
+A column file stores the values of the fields whose IDs are in its
`field_ids`. A list or map is stored as a whole, including every field nested
in it. For a struct, the column file stores whether the struct is null. Fields
nested in the struct are stored only if their IDs are also in `field_ids`. A
column file may also store fields whose IDs are not in its `field_ids`. Readers
must ignore these fields.
+
+When a column file is added, each ID in its `field_ids` must identify a field
valid for its base file. For a data file, fields in the current table schema
and `_last_updated_sequence_number` are valid. For a leaf manifest, fields in
the `tracked_file` struct are valid. Fields nested in a list or map are not
valid.
+
+#### Writing Column Files
+
+A column file stores a value for every row of its base file, including rows
whose values did not change. Values for deleted rows, and for fields in a null
struct, should be null. A column file may write a field as optional even if
that field is required in the table schema.
+
+An added column file must follow these rules for `field_ids`. A field must be
in `field_ids` if a rule requires it for at least one row where every struct
that the field is nested in is non-null after the update.
+
+- When a field nested in a list or map changes, the list or map must be in
`field_ids`.
+- Each non-struct field whose value changes must be in `field_ids`.
+- Each struct that changes between null and non-null must be in `field_ids`.
+- If a struct changes from null to non-null, each field directly nested in it
must be in `field_ids`, even if the field is null.
+
+See [Column file example](#column-file-example).
+
+When a writer adds column files to a data file, the writer must produce a
MODIFIED entry and a REPLACED entry. The writer must follow [Updating Tracking
Metadata](#updating-tracking-metadata) for those entries. The writer may write
the two entries to different manifests.
+
+When a writer adds column files to a leaf manifest, the writer must produce a
MODIFIED entry and a REPLACED entry in the root manifest. The REPLACED entry
must be a full entry in the root manifest. The writer must not replace the
previous leaf-manifest entry with a manifest deletion vector.
+
+The `tracking` fields of the MODIFIED entry are set as described in [Updating
Tracking Metadata](#updating-tracking-metadata) and [Sequence Number
Inheritance](#sequence-number-inheritance), and `tracking.snapshot_id` must be
the previous entry's (inherited) snapshot ID. The fields of the MODIFIED entry
outside `tracking` must match the previous entry, except for the following:
+
+- `partition`:
+ - For a data file, an added column file's `field_ids` may have the ID of a
source field of the data file's partition spec, or of a struct that one is
nested in, only if all live rows share the same partition values after the
update. `partition` must be set to those values.
+- `sort_order_id`:
+ - For a data file, if an added column file lists a field in the data
file's sort order, or a struct that one is nested in, `sort_order_id` must be
null.
+- `content_stats`:
+ - Stats for a field listed in an added column file, and for its nested
fields, must be recomputed or omitted.
+ - For a leaf manifest, stats that summarize a listed field must be
recomputed or omitted.
+- `tight_stats`:
+ - If `tight_stats` is true, `content_stats` must be tight for the live
rows after the update (see [Stats Tightness](#stats-tightness)).
+- `deletion_vector`:
+ - For a data file, every delete that applied to it before the column files
were added must be written into the MODIFIED entry's `deletion_vector`.
+- `column_files`:
+ - Each field listed in an added column file must be removed from the
`field_ids` of the previous entry's column files. A column file with no
remaining `field_ids` must be removed.
+ - For a data file, one of the added column files must have the
`_last_updated_sequence_number` ID in its `field_ids`, so that rows changed by
the added column files get a new `_last_updated_sequence_number`. In that
column file:
+ - If the added column files change the row,
`_last_updated_sequence_number` must be set to `null` (so that the sequence
number of the snapshot that added the column files replaces the current value).
+ - If the added column files do not change the row and the row is not
deleted, `_last_updated_sequence_number` must be copied to the new column file.
A null value must be replaced with the data sequence number a reader of the
previous entry would assign.
+
+#### Reading Column Files
+
+Each projected field, including each field nested in a projected field, must
be read from the column file whose `field_ids` has its ID. A field nested in a
list or map is read with that list or map. Other projected fields are read from
the base file, using [Column Projection](#column-projection) for a data file. A
column file of a data file is also read using [Column
Projection](#column-projection).
+
+If a struct is null in a row, every field nested in it is null in that row,
including a field read from a column file.
+
+A column file and its base file are aligned by row position before the
attached deletion vector is applied. The scan schema must be used for field
types and nullability. Null values in a column file may represent deleted rows.
+
+#### Column file example
+
+This example demonstrates which field IDs a column file must have in its
`field_ids`, using this schema and one row of a data file. Unless a row says
otherwise, that row in the base file is `customer = {name: …, address: {street:
'1 Elm', city: 'Austin'}}`. Column files are named `column_file_1`,
`column_file_2`, and so on. The `_last_updated_sequence_number` ID that each
update also adds to `field_ids` is left out.
+
+```
+2: optional struct customer {
+ 3: optional string name;
+ 4: optional struct address {
+ 5: optional string street;
+ 6: optional string city;
+ }
+}
+8: optional list<string> tags; // element field ID 9
+```
+
+| Row before the update | Update | Valid column files | Invalid column files |
Notes |
Review Comment:
I think this needs to be broken into smaller examples with data and not
changes or updates.
The spec should not state how you can go from one state to another, it
defines what valid states are.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]