diff --git a/VariantEncoding.md b/VariantEncoding.md new file mode 100644 index 000000000..1eac3bcbe --- /dev/null +++ b/VariantEncoding.md @@ -0,0 +1,429 @@ + + +# Variant Binary Encoding + +> [!IMPORTANT] +> **This specification is still under active development, and has not been formally adopted.** + +A Variant represents a type that contain one of: +- Primitive: A type and corresponding value (e.g. INT, STRING) +- Array: An ordered list of Variant values +- Object: An unordered collection of string/Variant pairs (i.e. key/value pairs). An object may not contain duplicate keys. + +A Variant is encoded with 2 binary values, the [value](#value-encoding) and the [metadata](#metadata-encoding). + +There are a fixed number of allowed primitive types, provided in the table below. +These represent a commonly supported subset of the [logical types](https://github.com/apache/parquet-format/blob/master/LogicalTypes.md) allowed by the Parquet format. + +The Variant Binary Encoding allows representation of semi-structured data (e.g. JSON) in a form that can be efficiently queried by path. +The design is intended to allow efficient access to nested data even in the presence of very wide or deep structures. + +Another motivation for the representation is that (aside from metadata) each nested Variant value is contiguous and self-contained. +For example, in a Variant containing an Array of Variant values, the representation of an inner Variant value, when paired with the metadata of the full variant, is itself a valid Variant. + +This document describes the Variant Binary Encoding scheme. +[VariantShredding.md](VariantShredding.md) describes the details of the Variant shredding scheme. + +# Variant in Parquet +A Variant value in Parquet is represented by a group with 2 fields, named `value` and `metadata`. +Both fields `value` and `metadata` are of type `binary`, and cannot be `null`. + +# Metadata encoding + +The encoded metadata always starts with a header byte. +``` + 7 6 5 4 3 0 + +-------+---+---+---------------+ +header | | | | version | + +-------+---+---+---------------+ + ^ ^ + | +-- sorted_strings + +-- offset_size_minus_one +``` +The `version` is a 4-bit value that must always contain the value `1`. +`sorted_strings` is a 1-bit value indicating whether dictionary strings are sorted and unique. +`offset_size_minus_one` is a 2-bit value providing the number of bytes per dictionary size and offset field. +The actual number of bytes, `offset_size`, is `offset_size_minus_one + 1`. + +The entire metadata is encoded as the following diagram shows: +``` + 7 0 + +-----------------------+ +metadata | header | + +-----------------------+ + | | + : dictionary_size : <-- little-endian, `offset_size` bytes + | | + +-----------------------+ + | | + : offset : <-- little-endian, `offset_size` bytes + | | + +-----------------------+ + : + +-----------------------+ + | | + : offset : <-- little-endian, `offset_size` bytes + | | (`dictionary_size + 1` offsets) + +-----------------------+ + | | + : bytes : + | | + +-----------------------+ +``` + +The metadata is encoded first with the `header` byte, then `dictionary_size` which is a little-endian value of `offset_size` bytes, and represents the number of string values in the dictionary. +Next, is an `offset` list, which contains `dictionary_size + 1` values. +Each `offset` is a little-endian value of `offset_size` bytes, and represents the starting byte offset of the i-th string in `bytes`. +The first `offset` value will always be `0`, and the last `offset` value will always be the total length of `bytes`. +The last part of the metadata is `bytes`, which stores all the string values in the dictionary. + +## Metadata encoding grammar + +The grammar for encoded metadata is as follows + +``` +metadata:
+header: 1 byte ( | << 4 | ( << 6)) +version: a 4-bit version ID. Currently, must always contain the value 1 +sorted_strings: a 1-bit value indicating whether metadata strings are sorted +offset_size_minus_one: 2-bit value providing the number of bytes per dictionary size and offset field. +dictionary_size: `offset_size` bytes. little-endian value indicating the number of strings in the dictionary +dictionary: * +offset: `offset_size` bytes. little-endian value indicating the starting position of the ith string in `bytes`. The list should contain `dictionary_size + 1` values, where the last value is the total length of `bytes`. +bytes: dictionary string values +``` + +Notes: +- Offsets are relative to the start of the `bytes` array. +- The length of the ith string can be computed as `offset[i+1] - offset[i]`. +- The offset of the first string is always equal to 0 and is therefore redundant. It is included in the spec to simplify in-memory-processing. +- `offset_size_minus_one` indicates the number of bytes per `dictionary_size` and `offset` entry. I.e. a value of 0 indicates 1-byte offsets, 1 indicates 2-byte offsets, 2 indicates 3 byte offsets and 3 indicates 4-byte offsets. +- If `sorted_strings` is set to 1, strings in the dictionary must be unique and sorted in lexicographic order. If the value is set to 0, readers may not make any assumptions about string order or uniqueness. + + +# Value encoding + +The entire encoded Variant value includes the `value_metadata` byte, and then 0 or more bytes for the `val`. +``` + 7 2 1 0 + +------------------------------------+------------+ +value | value_header | basic_type | + +------------------------------------+------------+ + | | + : value_data : <-- 0 or more bytes + | | + +-------------------------------------------------+ +``` +## Basic Type + +The `basic_type` is 2-bit value that represents which basic type the Variant value is. +The [basic types table](#encoding-types) shows what each value represents. + +## Value Header + +The `value_header` is a 6-bit value that contains more information about the type, and the format depends on the `basic_type`. + +### Value Header for Primitive type (`basic_type`=0) + +When `basic_type` is `0`, `value_header` is a 6-bit `primitive_header`. +The [primitive types table](#encoding-types) shows what each value represents. +``` + 5 0 + +-----------------------+ +value_header | primitive_header | + +-----------------------+ +``` + +### Value Header for Short string (`basic_type`=1) + +When `basic_type` is `1`, `value_header` is a 6-bit `short_string_header`. +``` + 5 0 + +-----------------------+ +value_header | short_string_header | + +-----------------------+ +``` +The `short_string_header` value is the length of the string. + +### Value Header for Object (`basic_type`=2) + +When `basic_type` is `2`, `value_header` is made up of `field_offset_size_minus_one`, `field_id_size_minus_one`, and `is_large`. +``` + 5 4 3 2 1 0 + +---+---+-------+-------+ +value_header | | | | | + +---+---+-------+-------+ + ^ ^ ^ + | | +-- field_offset_size_minus_one + | +-- field_id_size_minus_one + +-- is_large +``` +`field_offset_size_minus_one` and `field_id_size_minus_one` are 2-bit values that represent the number of bytes used to encode the field offsets and field ids. +The actual number of bytes is computed as `field_offset_size_minus_one + 1` and `field_id_size_minus_one + 1`. +`is_large` is a 1-bit value that indicates how many bytes are used to encode the number of elements. +If `is_large` is `0`, 1 byte is used, and if `is_large` is `1`, 4 bytes are used. + +### Value Header for Array (`basic_type`=3) + +When `basic_type` is `3`, `value_header` is made up of `field_offset_size_minus_one`, and `is_large`. +``` + 5 3 2 1 0 + +-----------+---+-------+ +value_header | | | | + +-----------+---+-------+ + ^ ^ + | +-- field_offset_size_minus_one + +-- is_large +``` +`field_offset_size_minus_one` is a 2-bit value that represents the number of bytes used to encode the field offset. +The actual number of bytes is computed as `field_offset_size_minus_one + 1`. +`is_large` is a 1-bit value that indicates how many bytes are used to encode the number of elements. +If `is_large` is `0`, 1 byte is used, and if `is_large` is `1`, 4 bytes are used. + +## Value Data + +The `value_data` encoding format depends on the type specified by `value_metadata`. +For some types, the `value_data` will be 0-bytes. + +### Value Data for Primitive type (`basic_type`=0) + +When `basic_type` is `0`, `value_data` depends on the `primitive_header` value. +The [primitive types table](#encoding-types) shows the encoding format for each primitive type. + +### Value Data for Short string (`basic_type`=1) + +When `basic_type` is `1`, `value_data` is the sequence of bytes that represents the string. + +### Value Data for Object (`basic_type`=2) + +When `basic_type` is `2`, `value_data` encodes an object. +The encoding format is shown in the following diagram: +``` + 7 0 + +-----------------------+ +object value_data | | + : num_elements : <-- little-endian, 1 or 4 bytes + | | + +-----------------------+ + | | + : field_id : <-- little-endian, `field_id_size` bytes + | | + +-----------------------+ + : + +-----------------------+ + | | + : field_id : <-- little-endian, `field_id_size` bytes + | | (`num_elements` field_ids) + +-----------------------+ + | | + : field_offset : <-- little-endian, `field_offset_size` bytes + | | + +-----------------------+ + : + +-----------------------+ + | | + : field_offset : <-- little-endian, `field_offset_size` bytes + | | (`num_elements + 1` field_offsets) + +-----------------------+ + | | + : value : + | | + +-----------------------+ + : + +-----------------------+ + | | + : value : <-- (`num_elements` values) + | | + +-----------------------+ +``` +An object `value_data` begins with `num_elements`, a 1-byte or 4-byte little-endian value, representing the number of elements in the object. +The size in bytes of `num_elements` is indicated by `is_large` in the `value_header`. +Next, is a list of `field_id` values. +There are `num_elements` number of entries and each `field_id` is a little-endian value of `field_id_size` bytes. +A `field_id` is an index into the dictionary in the metadata. +The `field_id` list is followed by a `field_offset` list. +There are `num_elements + 1` number of entries and each `field_offset` is a little-endian value of `field_offset_size` bytes. +A `field_offset` represents the byte offset (relative to the first byte of the first `value`) where the i-th `value` starts. +The last `field_offset` points to the byte after the end of the last `value`. +The `field_offset` list is followed by the `value` list. +There are `num_elements` number of `value` entries and each `value` is an encoded Variant value. +For the i-th key-value pair of the object, the key is the metadata dictionary entry indexed by the i-th `field_id`, and the value is the Variant `value` starting from the i-th `field_offset` byte offset. + +The field ids and field offsets must be in lexicographical order of the corresponding field names in the metadata dictionary. +However, the actual `value` entries do not need to be in any particular order. +This implies that the `field_offset` values may not be monotonically increasing. +For example, for the following object: +``` +{ + "c": 3, + "b": 2, + "a": 1 +} +``` +The `field_id` list must be `[, , ]`, in lexicographical order. +The `field_offset` list must be `[, , , ]`. +The `value` list can be in any order. + +### Value Data for Array (`basic_type`=3) + +When `basic_type` is `3`, `value_data` encodes an array. The encoding format is shown in the following diagram: +``` + 7 0 + +-----------------------+ +array value_data | | + : num_elements : <-- little-endian, 1 or 4 bytes + | | + +-----------------------+ + | | + : field_offset : <-- little-endian, `field_offset_size` bytes + | | + +-----------------------+ + : + +-----------------------+ + | | + : field_offset : <-- little-endian, `field_offset_size` bytes + | | (`num_elements + 1` field_offsets) + +-----------------------+ + | | + : value : + | | + +-----------------------+ + : + +-----------------------+ + | | + : value : <-- (`num_elements` values) + | | + +-----------------------+ +``` +An array `value_data` begins with `num_elements`, a 1-byte or 4-byte little-endian value, representing the number of elements in the array. +The size in bytes of `num_elements` is indicated by `is_large` in the `value_header`. +Next, is a `field_offset` list. +There are `num_elements + 1` number of entries and each `field_offset` is a little-endian value of `field_offset_size` bytes. +A `field_offset` represents the byte offset (relative to the first byte of the first `value`) where the i-th `value` starts. +The last `field_offset` points to the byte after the last byte of the last `value`. +The `field_offset` list is followed by the `value` list. +There are `num_elements` number of `value` entries and each `value` is an encoded Variant value. +For the i-th array entry, the value is the Variant `value` starting from the i-th `field_offset` byte offset. + +## Value encoding grammar + +The grammar for an encoded value is: + +``` +value: ? +value_metadata: 1 byte ( | ( << 2)) +basic_type: ID from Basic Type table. must be a corresponding variation +value_header: | | | +primitive_header: ID from Primitive Type table. must be a corresponding variation of +short_string_header: unsigned string length in bytes from 0 to 63 +object_header: (is_large << 4 | field_id_size_minus_one << 2 | field_offset_size_minus_one) +array_header: (is_large << 2 | field_offset_size_minus_one) +value_data: | | | +primitive_val: see table for binary representation +short_string_val: bytes +object_val: * * +array_val: * +num_elements: a 1 or 4 byte little-endian value (depending on is_large in /) +field_id: a 1, 2, 3 or 4 byte little-endian value (depending on field_id_size_minus_one in ), indexing into the dictionary +field_offset: a 1, 2, 3 or 4 byte little-endian value (depending on field_offset_size_minus_one in /), providing the offset in bytes within fields +fields: * +``` + +Each `value_data` must correspond to the type defined by `value_metadata`. Boolean and null types do not have a corresponding `value_data`, since their type defines their value. + +Each `array_val` and `object_val` must contain exactly `num_elements + 1` values for `field_offset`. +The last entry is the offset that is one byte past the last field (i.e. the total size of all fields in bytes). +All offsets are relative to the first byte of the first field in the object/array. + +`field_id_size_minus_one` and `field_offset_size_minus_one` indicate the number of bytes per field ID/offset. +For example, a value of 0 indicates 1-byte IDs, 1 indicates 2-byte IDs, 2 indicates 3 byte IDs and 3 indicates 4-byte IDs. +The `is_large` flag for arrays and objects is used to indicate whether the number of elements is indicated using a one or four byte value. +When more than 255 elements are present, `is_large` must be set to true. +It is valid for an implementation to use a larger value than necessary for any of these fields (e.g. `is_large` may be true for an object with less than 256 elements). + +The "short string" basic type may be used as an optimization to fold string length into the type byte for strings less than 64 bytes. +It is semantically identical to the "string" primitive type. + +The Decimal type contains a scale, but no precision. The implied precision of a decimal value is `floor(log_10(val)) + 1`. + +# Encoding types + +| Basic Type | ID | Description | +|--------------|-----|---------------------------------------------------| +| Primitive | `0` | One of the primitive types | +| Short string | `1` | A string with a length less than 64 bytes | +| Object | `2` | A collection of (string-key, variant-value) pairs | +| Array | `3` | An ordered sequence of variant values | + +| Logical Type | Physical Type | Type ID | Equivalent Parquet Type | Binary format | +|----------------------|-----------------------------|---------|-----------------------------|---------------------------------------------------------------------------------------------------------------------| +| NullType | null | `0` | any | none | +| Boolean | boolean (True) | `1` | BOOLEAN | none | +| Boolean | boolean (False) | `2` | BOOLEAN | none | +| Exact Numeric | int8 | `3` | INT(8, signed) | 1 byte | +| Exact Numeric | int16 | `4` | INT(16, signed) | 2 byte little-endian | +| Exact Numeric | int32 | `5` | INT(32, signed) | 4 byte little-endian | +| Exact Numeric | int64 | `6` | INT(64, signed) | 8 byte little-endian | +| Double | double | `7` | DOUBLE | IEEE little-endian | +| Exact Numeric | decimal4 | `8` | DECIMAL(precision, scale) | 1 byte scale in range [0, 38], followed by little-endian unscaled value (see decimal table) | +| Exact Numeric | decimal8 | `9` | DECIMAL(precision, scale) | 1 byte scale in range [0, 38], followed by little-endian unscaled value (see decimal table) | +| Exact Numeric | decimal16 | `10` | DECIMAL(precision, scale) | 1 byte scale in range [0, 38], followed by little-endian unscaled value (see decimal table) | +| Date | date | `11` | DATE | 4 byte little-endian | +| Timestamp | timestamp | `12` | TIMESTAMP(true, MICROS) | 8-byte little-endian | +| TimestampNTZ | timestamp without time zone | `13` | TIMESTAMP(false, MICROS) | 8-byte little-endian | +| Float | float | `14` | FLOAT | IEEE little-endian | +| Binary | binary | `15` | BINARY | 4 byte little-endian size, followed by bytes | +| String | string | `16` | STRING | 4 byte little-endian size, followed by UTF-8 encoded bytes | + +| Decimal Precision | Decimal value type | +|-----------------------|--------------------| +| 1 <= precision <= 9 | int32 | +| 10 <= precision <= 18 | int64 | +| 18 <= precision <= 38 | int128 | +| > 38 | Not supported | + +The *Logical Type* column indicates logical equivalence of physically encoded types. +For example, a user expression operating on a string value containing "hello" should behave the same, whether it is encoded with the short string optimization, or long string encoding. +Similarly, user expressions operating on an *int8* value of 1 should behave the same as a decimal16 with scale 2 and unscaled value 100. + +# Field ID order and uniqueness + +For objects, field IDs and offsets must be listed in the order of the corresponding field names, sorted lexicographically. +Note that the fields themselves are not required to follow this order. +As a result, offsets will not necessarily be listed in ascending order. + +An implementation may rely on this field ID order in searching for field names. +E.g. a binary search on field IDs (combined with metadata lookups) may be used to find a field with a given field. + +Field names are case-sensitive. +Field names are required to be unique for each object. +It is an error for an object to contain two fields with the same name, whether or not they have distinct dictionary IDs. + +# Versions and extensions + +An implementation is not expected to parse a Variant value whose metadata version is higher than the version supported by the implementation. +However, new types may be added to the specification without incrementing the version ID. +In such a situation, an implementation should be able to read the rest of the Variant value if desired. + +# Shredding + +A single Variant object may have poor read performance when only a small subset of fields are needed. +A better approach is to create separate columns for individual fields, referred to as shredding or subcolumnarization. +[VariantShredding.md](VariantShredding.md) describes the Variant shredding specification in Parquet. diff --git a/VariantShredding.md b/VariantShredding.md new file mode 100644 index 000000000..51160a9bc --- /dev/null +++ b/VariantShredding.md @@ -0,0 +1,300 @@ + + +# Variant Shredding + +> [!IMPORTANT] +> **This specification is still under active development, and has not been formally adopted.** + +The Variant type is designed to store and process semi-structured data efficiently, even with heterogeneous values. +Query engines encode each Variant value in a self-describing format, and store it as a group containing `value` and `metadata` binary fields in Parquet. +Since data is often partially homogenous, it can be beneficial to extract certain fields into separate Parquet columns to further improve performance. +We refer to this process as **shredding**. +Each Parquet file remains fully self-describing, with no additional metadata required to read or fully reconstruct the Variant data from the file. +Combining shredding with a binary residual provides the flexibility to represent complex, evolving data with an unbounded number of unique fields while limiting the size of file schemas, and retaining the performance benefits of a columnar format. + +This document focuses on the shredding semantics, Parquet representation, implications for readers and writers, as well as the Variant reconstruction. +For now, it does not discuss which fields to shred, user-facing API changes, or any engine-specific considerations like how to use shredded columns. +The approach builds upon the [Variant Binary Encoding](VariantEncoding.md), and leverages the existing Parquet specification. + +At a high level, we replace the `value` field of the Variant Parquet group with one or more fields called `object`, `array`, `typed_value`, and `variant_value`. +These represent a fixed schema suitable for constructing the full Variant value for each row. + +Shredding allows a query engine to reap the full benefits of Parquet's columnar representation, such as more compact data encoding, min/max statistics for data skipping, and I/O and CPU savings from pruning unnecessary fields not accessed by a query (including the non-shredded Variant binary data). +Without shredding, any query that accesses a Variant column must fetch all bytes of the full binary buffer. +With shredding, we can get nearly equivalent performance as in a relational (scalar) data model. + +For example, `select variant_get(variant_col, ‘$.field1.inner_field2’, ‘string’) from tbl` only needs to access `inner_field2`, and the file scan could avoid fetching the rest of the Variant value if this field was shredded into a separate column in the Parquet schema. +Similarly, for the query `select * from tbl where variant_get(variant_col, ‘$.id’, ‘integer’) = 123`, the scan could first decode the shredded `id` column, and only fetch/decode the full Variant value for rows that pass the filter. + +# Parquet Example + +Consider the following Parquet schema together with how Variant values might be mapped to it. +Notice that we represent each shredded field in `object` as a group of two fields, `typed_value` and `variant_value`. +We extract all homogenous data items of a certain path into `typed_value`, and set aside incompatible data items in `variant_value`. +Intuitively, incompatibilities within the same path may occur because we store the shredding schema per Parquet file, and each file can contain several row groups. +Selecting a type for each field that is acceptable for all rows would be impractical because it would require buffering the contents of an entire file before writing. + +Typically, the expectation is that `variant_value` exists at every level as an option, along with one of `object`, `array` or `typed_value`. +If the actual Variant value contains a type that does not match the provided schema, it is stored in `variant_value`. +An `variant_value` may also be populated if an object can be partially represented: any fields that are present in the schema must be written to those fields, and any missing fields are written to `variant_value`. + +The `metadata` column is unchanged from its unshredded representation, and may be referenced in `variant_value` fields in the shredded data. + +``` +optional group variant_col { + required binary metadata; + optional binary variant_value; + optional group object { + optional group a { + optional binary variant_value; + optional int64 typed_value; + } + optional group b { + optional binary variant_value; + optional group object { + optional group c { + optional binary variant_value; + optional binary typed_value (STRING); + } + } + } + } +} +``` + +| Variant Value | Top-level variant_value | b.variant_value | a.typed_value | a.variant_value | b.object.c.typed_value | b.object.c.variant_value | Notes | +|---------------|-------------------------|-----------------|---------------|-----------------|------------------------|--------------------------|-------| +| {a: 123, b: {c: “hello”}} | null | null | 123 | null | hello | null | All values shredded | +| {a: 1.23, b: {c: “123”}} | null | null | null | 1.23 | 123 | null | a is not an integer | +| {a: 123, b: {c: null}} | null | null | null | 123 | null | null | b.object.c set to non-null to indicate VariantNull | +| {a: 123, b: {} | null | null | null | 123 | null | null | b.object.c set to null, to indicate that c is missing | +| {a: 123, d: 456} | {d: 456} | null | 123 | null | null | null | Extra field d is stored as variant_value | +| [{a: 1, b: {c: 2}}, {a: 3, b: {c: 4}}] | [{a: 1, b: {c: 2}}, {a: 3, b: {c: 4}}] | null | null | null | null | null | Not an object | + +# Parquet Layout + +The `array` and `object` fields represent Variant array and object types, respectively. +Arrays must use the three-level list structure described in https://github.com/apache/parquet-format/blob/master/LogicalTypes.md. + +An `object` field must be a group. +Each field name of this inner group corresponds to the Variant value's object field name. +Each inner field's type is a recursively shredded variant value: that is, the fields of each object field must be one or more of `object`, `array`, `typed_value` or `variant_value`. + +Similarly the elements of an `array` must be a group containing one or more of `object`, `array`, `typed_value` or `variant_value`. + +Each leaf in the schema can store an arbitrary Variant value. +It contains an `variant_value` binary field and a `typed_value` field. +If non-null, `variant_value` represents the value stored as a Variant binary. +The `typed_value` field may be any type that has a corresponding Variant type. +For each value in the data, at most one of the `typed_value` and `variant_value` may be non-null. +A writer may omit either field, which is equivalent to all rows being null. + +Dictionary IDs in a `variant_value` field refer to entries in the top-level `metadata` field. + +For an `object`, a null field means that the field does not exist in the reconstructed Variant object. +All elements of an `array` must be non-null, since array elements cannote be missing. + +| typed_value | variant_value | Meaning | +|-------------|----------------|---------| +| null | null | Field is Variant Null (not missing) in the reconstructed Variant. | +| null | non-null | Field may be any type in the reconstructed Variant. | +| non-null | null | Field has this column’s type in the reconstructed Variant. | +| non-null | non-null | Invalid | + +The `typed_value` may be absent from the Parquet schema for any field, which is equivalent to its value being always null (in which case the shredded field is always stored as a Variant binary). +By the same token, `variant_value` may be absent, which is equivalent to their value being always null (in which case the field will always have the value Null or have the type of the `typed_value` column). + +# Unshredded values + +If all values can be represented at a given level by whichever of `object`, `array`, or `typed_value` is present, `variant_value` is set to null. + +If a value cannot be represented by whichever of `object`, `array`, or `typed_value` is present in the schema, then it is stored in `variant_value`, and the other fields are set to null. +In the Parquet example above, if field `a` was an object or array, or a non-integer scalar, it would be stored in `variant_value`. + +If a value is an object, and the `object` field is present but does not contain all of the fields in the value, then any remaining fields are stored in an object in `variant_value`. +In the Parquet example above, if field `b` was an object of the form `{"c": 1, "d": 2}"`, then the object `{"d": 2}` would be stored in `variant_value`, and the `c` field would be shredded recursively under `object.c`. + +Note that an array is always fully shredded if there is an `array` field, so the above consideration for `object` is not relevant for arrays: only one of `array` or `variant_value` may be non-null at a given level. + +# Using variant_value vs. typed_value + +In general, it is desirable to store values in the `typed_value` field rather than the `variant_value` whenever possible. +This will typically improve encoding efficiency, and allow the use of Parquet statistics to filter at the row group or page level. +In the best case, the `variant_value` fields are all null and the engine does not need to read them (or it can omit them from the schema on write entirely). +There are two main motivations for including the `variant_value` column: + +1) In a case where there are rare type mismatches (for example, a numeric field with rare strings like “n/a”), we allow the field to be shredded, which could still be a significant performance benefit compared to fetching and decoding the full value/metadata binary. +2) Since there is a single schema per file, there would be no easy way to recover from a type mismatch encountered late in a file write. Parquet files can be large, and buffering all file data before starting to write could be expensive. Including a variant column for every field guarantees we can adhere to the requested shredding schema. + +# Data Skipping + +Shredded columns are expected to store statistics in the same format as a normal Parquet column. +In general, the engine can only skip a row group or page if all rows in the `variant_value` field are null, since it is possible for a `variant_get` expression to successfully cast a value from the `variant_value` to the target type. +For example, if `typed_value` is of type `int64`, then the string “123” might be contained in `variant_value`, which would not be reflected in statistics, but could be retained by a filter like `where variant_get(col, “$.field”, “long”) = 123`. +If `variant_value` is all-null, then the engine can prune pages or row groups based on `typed_value`. +This specification is not strict about what values may be stored in `variant_value` rather than `typed_value`, so it is not safe to skip rows based on `typed_value` unless the corresponding `variant_value` column is all-null, or the engine has specific knowledge of the behavior of the writer that produced the shredded data. + +# Shredding Semantics + +Reconstruction of Variant value from a shredded representation is not expected to produce a bit-for-bit identical binary to the original unshredded value. +For example, the order of fields in the binary may change, as may the physical representation of scalar values. + +In particular, the [Variant Binary Encoding](VariantEncoding.md) considers all integer and decimal representations to represent a single logical type. +As a result, it is valid to shred a decimal into a decimal column with a different scale, or to shred an integer as a decimal, as long as no numeric precision is lost. +For example, it would be valid to write the value 123 to a Decimal(9, 2) column, but the value 1.234 would need to be written to the **variant_value** column. +When reconstructing, it would be valid for a reader to reconstruct 123 as an integer, or as a Decimal(9, 2). +Engines should not depend on the physical type of a Variant value, only the logical type. + +On the other hand, shredding as a different logical type is not allowed. +For example, the integer value 123 could not be shredded to a string `typed_value` column as the string "123", since that would lose type information. +It would need to be written to the `variant_value` column. + +# Reconstructing a Variant + +It is possible to recover a full Variant value using a recursive algorithm, where the initial call is to `ConstructVariant` with the top-level fields, which are assumed to be null if they are not present in the schema. + +``` +# Constructs a Variant from `variant_value`, `object`, `array` and `typed_value`. +# Only one of object, array and typed_value may be non-null. +def ConstructVariant(variant_value, object, array, typed_value): + if object is null and array is null and typed_value is null and variant_value is null: return VariantNull + if object is not null: + return ConstructObject(variant_value, object) + elif array is not null: + return ConstructArray(array) + elif typed_value is not null: + return cast(typed_value as Variant) + else: + variant_value + +# Construct an object from an `object` group, and a (possibly null) Variant variant_value +def ConstructObject(variant_value, object): + # If variant_value is present and is not an Object, then the result is ambiguous. + assert(variant_value is null or is_object(variant_value)) + # Null fields in the object are missing from the reconstructed Variant. + nonnull_object_fields = object.fields.filter(field -> field is not null) + all_keys = Union(variant_value.keys, non_null_object_fields) + return VariantObject(all_keys.map { key -> + if key in object: (key, ConstructVariant(object[key].variant_value, object[key].object, object[key].array, object[key].typed_value)) + else: (key, variant_value[key]) + }) + +def ConstructArray(array): + newVariantArray = VariantArray() + for i in range(array.size): + newVariantArray.append(ConstructVariant(array[i].variant_value, array[i].object, array[i].array, array[i].typed_value) +``` + +# Nested Parquet Example + +This section describes a more deeply nested example, using a top-level array as the shredding type. + +Below is a sample of JSON that would be fully shredded in this example. +It contains an array of objects, containing an `a` field shredded as an array, and a `b` field shredded as an integer. + +``` +[ + { + "a": [1, 2, 3], + "b": 100 + }, + { + "a": [4, 5, 6], + "b": 200 + } +] +``` + + +The corresponding Parquet schema with “a” and “b” as leaf types is: + +``` +optional group variant_col { + required binary metadata; + optional binary variant_value; + optional group array (LIST) { + repeated group list { + optional group element { + optional binary variant_value; + optional group object { + optional group a { + optional binary variant_value; + optional group array (LIST) { + repeated group list { + optional group element { + optional int64 typed_value; + optional binary variant_value; + } + } + } + } + optional group b { + optional int64 typed_value; + optional binary variant_value; + } + } + } + } + } +} +``` + +In the above example schema, if “a” is an array containing a mix of integer and non-integer values, the engine will shred individual elements appropriately into either `typed_value` or `variant_value`. +If the top-level Variant is not an array (for example, an object), the engine cannot shred the value and it will store it in the top-level `variant_value`. +Similarly, if "a" is not an array, it will be stored in the `variant_value` under "a". + +Consider the following example: + +``` +[ + { + "a": [1, 2, 3], + "b": 100, + “c”: “unexpected” + }, + { + "a": [4, 5, 6], + "b": 200 + }, + “not an object” +] +``` + +The second array element can be fully shredded, but the first and third cannot be. The contents of `variant_col.array[*].variant_value` would be as follows: + +``` +[ + { “c”: “unexpected” }, + NULL, + “not an object” +] +``` + +# Backward and forward compatibility + +Shredding is an optional feature of Variant, and readers must continue to be able to read a group containing only a `value` and `metadata` field. + +Any fields in the same group as `typed_value`/`variant_value` that start with `_` (underscore) can be ignored. +This is intended to allow future backwards-compatible extensions. +In particular, the field names `_metadata_key_paths` and any name starting with `_spark` are reserved, and should not be used by other implementations. +Any extra field names that do not start with an underscore should be assumed to be backwards incompatible, and readers should fail when reading such a schema. + +Engines without shredding support are not expected to be able to read Parquet files that use shredding. +Since different files may contain conflicting schemas (e.g. a `typed_value` column with incompatible types in two files), it may not be possible to infer or specify a single schema that would allow all Parquet files for a table to be read.