Dataset Viewer
Auto-converted to Parquet Duplicate
scenario_id
string
title
string
category
string
description
string
notes
list
tags
list
row_identity
list
version_count
int64
version_ids
list
row_counts
list
mutation_count
int64
mutation_kinds
list
paths
list
value_change_relations
list
schema_changed
bool
values_changed
bool
parquet_schema_changed
bool
base_file
string
final_file
string
base_schema_json
string
final_schema_json
string
transitions_json
string
base_fingerprint
string
final_fingerprint
string
manifest_path
string
add_all_null_column
Add a typed all-null column, then backfill it
additive
v1 adds a string column in which every value is null. v2 backfills it without any schema change. The column is typed string in both v1 and v2; only its values differ.
[ "The v1 column is declared string, not the Arrow null type. A column that is inferred as the null type from all-null data is a different case: see null_type_to_string." ]
[ "add-field", "all-null", "backfill", "multi-version" ]
[ "id" ]
3
[ "v0", "v1", "v2" ]
[ 6, 6, 6 ]
1
[ "add_field" ]
[ "referrer" ]
[ "nulls_filled" ]
true
true
true
fixtures/add_all_null_column/v0.parquet
fixtures/add_all_null_column/v2.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"referrer","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":true,"path":"referrer","type":"string"}],"to":"v1"},{"from":"v1"...
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
sha256:8765da5f8f45127f829ad9538ef86cfee69ec7e05752d7af32209e19b15f51c2
fixtures/add_all_null_column/scenario.json
add_field_in_list_of_struct
Add a field to the structs inside a list
structural
Each element of a list<struct> column gains a nullable float64 field discount. Rows include a null list, an empty list and a list with a null element. Existing element values are unchanged.
[]
[ "list", "struct", "nested", "add-field", "nested-nulls" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "add_field" ]
[ "line_items[].discount" ]
[]
true
false
true
fixtures/add_field_in_list_of_struct/v0.parquet
fixtures/add_field_in_list_of_struct/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"line_items","nullable":true,"type":"list<struct<sku: string, qty: int32>>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"line_items","nullable":true,"type":"list<struct<sku: string, qty: int32, discount: double>>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":true,"path":"line_items[].discount","type":"double"}],"to":"v1"}...
sha256:70cac456c966e83ece8cbfa7e878625a2f2551db19687cf90590f2c2439e9e08
sha256:529bf08d7640137b7763e571a81bc2f48871c8d089443f3b4f8f5fe8a0f1e564
fixtures/add_field_in_list_of_struct/scenario.json
add_nested_field
Add a field inside a struct
structural
A nullable string field postcode is appended to the address struct. Rows include a null struct, a struct whose street is null, and a struct of empty strings; a null struct stays null. Existing nested values are unchanged.
[]
[ "struct", "nested", "add-field", "nested-nulls" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "add_field" ]
[ "address.postcode" ]
[]
true
false
true
fixtures/add_nested_field/v0.parquet
fixtures/add_nested_field/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"address","nullable":true,"type":"struct<street: string, city: string>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"address","nullable":true,"type":"struct<street: string, city: string, postcode: string>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":true,"path":"address.postcode","type":"string"}],"to":"v1"}]
sha256:fe2fd9399e594ce5ae78b4ba68f7dac7d891cef4ef9ab4eafc2742d1b7eb6cfc
sha256:ffc0803e00bf07b7da8ccb4f60df1121b9f160cf58c79b6a505b5450a2aeef28
fixtures/add_nested_field/scenario.json
add_nullable_int64_column
Add nullable int64 column
additive
A nullable int64 column is appended. Its values include zero, a negative value, nulls and both int64 extremes. Existing columns and rows are unchanged.
[]
[ "add-field", "integer", "nullable", "numeric-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "add_field" ]
[ "loyalty_points" ]
[]
true
false
true
fixtures/add_nullable_int64_column/v0.parquet
fixtures/add_nullable_int64_column/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"loyalty_points","nullable":true,"type":"int64"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":true,"path":"loyalty_points","type":"int64"}],"to":"v1"}]
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
sha256:0e2387711427d5d742df3b176e5b6190ad79df5438382c28e06e45ee9ff1233e
fixtures/add_nullable_int64_column/scenario.json
add_nullable_string_column
Add nullable string column
additive
A nullable string column is appended. Every existing column, value and row is unchanged. The new column holds both nulls and empty strings, which are different values.
[]
[ "add-field", "string", "nullable" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "add_field" ]
[ "customer_segment" ]
[]
true
false
true
fixtures/add_nullable_string_column/v0.parquet
fixtures/add_nullable_string_column/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"customer_segment","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":true,"path":"customer_segment","type":"string"}],"to":"v1"}]
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
sha256:3d0adf61168f53d1f9fcfbc45c5b056f18a1acdbfd57826e5dd6d970568ca47c
fixtures/add_nullable_string_column/scenario.json
add_required_column
Add non-nullable populated column
additive
A non-nullable string column is appended, with a value in every v1 row. v0 has no such column, so the v0 rows have no value for it. Existing columns and rows are unchanged.
[ "Read by column name together with v0, the v0 rows have no value for a column that v1 declares non-nullable. See remove_required_column for the opposite direction." ]
[ "add-field", "string", "non-nullable" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "add_field" ]
[ "country_code" ]
[]
true
false
true
fixtures/add_required_column/v0.parquet
fixtures/add_required_column/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"country_code","nullable":false,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"add_field","nullable":false,"path":"country_code","type":"string"}],"to":"v1"}]
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
sha256:e9c57949b192096eb9a46534c80fabe05061656018bccc1fc10bb57ae82616bb
fixtures/add_required_column/scenario.json
column_becomes_all_null
Populated column becomes entirely null
nullability
Every value of a nullable float64 column is null in v1. The schema is unchanged. v0 contains NaN, -0.0 and infinity, none of which is a null.
[ "v0 has exactly one null. Systems that count NaN as missing report two; that is a policy of those systems, not a property of the file." ]
[ "nullable", "all-null", "no-schema-change", "nan", "infinity" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
0
[]
[ "discount" ]
[ "nulled" ]
false
true
false
fixtures/column_becomes_all_null/v0.parquet
fixtures/column_becomes_all_null/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"discount","nullable":true,"type":"double"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"discount","nullable":true,"type":"double"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"discount","relation":"nulled"}]},"mutations":[],"to":"v1"}]
sha256:3ef9076f485c18b6a67b93cbbba562eb3554c37b29f734567bf15bd45da5d1ba
sha256:3977f21592cfa279f263a62b343471ef4a20347ffe6a6001cfab216f43d39751
fixtures/column_becomes_all_null/scenario.json
decimal_precision_increase
Increase decimal precision
numeric
A decimal128(9, 2) column becomes decimal128(18, 2). Scale is unchanged. Values include both extremes of the old precision and are unchanged.
[ "The Parquet fixed_len_byte_array width changes with the precision (4 to 8 bytes)." ]
[ "change-type", "decimal", "widening", "numeric-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "price" ]
[]
true
false
true
fixtures/decimal_precision_increase/v0.parquet
fixtures/decimal_precision_increase/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"price","nullable":true,"type":"decimal128(9, 2)"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"price","nullable":true,"type":"decimal128(18, 2)"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"decimal128(18, 2)","old_type":"decimal128(9, 2)","path":"pric...
sha256:82d7691f28b4dad15f955c446e3a3dde7ead4c38b122f5fdce33ccd26632f258
sha256:c7fe98a1141c4d570b4c786a20f29140f577dcc658ff87544952742adde45e7c
fixtures/decimal_precision_increase/scenario.json
decimal_scale_increase
Increase decimal scale at fixed precision
numeric
A decimal128(9, 2) column becomes decimal128(9, 4). Precision is unchanged, so the largest representable magnitude shrinks from 9999999.99 to 99999.9999. Every v0 value still fits and every value is numerically unchanged (1.20 and 1.2000 are the same number).
[ "The type change narrows the value range while widening the scale; a v0 value such as 100000.00 would not fit. None of the fixture values are affected." ]
[ "change-type", "decimal", "scale" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "price" ]
[]
true
false
true
fixtures/decimal_scale_increase/v0.parquet
fixtures/decimal_scale_increase/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"price","nullable":true,"type":"decimal128(9, 2)"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"price","nullable":true,"type":"decimal128(9, 4)"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"decimal128(9, 4)","old_type":"decimal128(9, 2)","path":"price...
sha256:6704cd50d57650f1d356aea7ab387d2b9b563c9981aaeba7796306ca527e6bd1
sha256:a79e0ee3fdef5f5be10b5e383e6270e8878eee342b012bdefdd6d4608cb4beca
fixtures/decimal_scale_increase/scenario.json
dictionary_reencoded
Dictionary contents change without any value change
representation
Both versions store the same dictionary<int8, string> column with the same decoded values. The stored dictionaries differ: v0 has a different order and an entry that no row uses; v1 has exactly the used values. Schema, rows and values are unchanged, so both versions have the same logical fingerprint.
[ "Tools that compare dictionary indices, category lists or category codes observe a change here. The decoded values do not change.", "The unused dictionary entry survives a PyArrow round trip; other readers may rebuild the dictionary on read and drop it." ]
[ "dictionary", "categorical", "encoding", "no-schema-change", "no-value-change" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
0
[]
[]
[]
false
false
false
fixtures/dictionary_reencoded/v0.parquet
fixtures/dictionary_reencoded/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"dictionary<values=string, indices=int8, ordered=0>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"dictionary<values=string, indices=int8, ordered=0>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[],"to":"v1"}]
sha256:6dfdbf0a9de0eac2a01b3d7a88cb06abe18009dfa4d033259c14f39d7bf16d9b
sha256:6dfdbf0a9de0eac2a01b3d7a88cb06abe18009dfa4d033259c14f39d7bf16d9b
fixtures/dictionary_reencoded/scenario.json
empty_then_populated
Empty typed file followed by a populated file
edge
v0 has a fully typed schema and zero rows, the Parquet counterpart of a header-only CSV. v1 has the identical schema and five rows whose values include NaN, both infinities, -0.0, an empty string, whitespace, the literal text NULL and a timestamp before the epoch.
[ "The row and value invariants quantify over the rows of v0, of which there are none, so rows_retained and common_values_preserved are vacuously true.", "v0 carries complete type information in its schema and none in its values, because it has no values." ]
[ "empty", "zero-rows", "no-schema-change", "nan", "infinity", "unicode" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 0, 5 ]
0
[]
[]
[]
false
false
false
fixtures/empty_then_populated/v0.parquet
fixtures/empty_then_populated/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"label","nullable":true,"type":"string"},{"name":"amount","nullable":true,"type":"double"},{"name":"observed_at","nullable":true,"type":"timestamp[us, tz=UTC]"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"label","nullable":true,"type":"string"},{"name":"amount","nullable":true,"type":"double"},{"name":"observed_at","nullable":true,"type":"timestamp[us, tz=UTC]"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":false,"row_identity_preserved":false,"rows_retained":true,"value_changes":[]},"mutations":[],"to":"v1"}]
sha256:6ec61b2ba1c3d3822c5a21ae88514cfe9517ae7c1b24b0a1d7d51444304ae6a6
sha256:4eef0003cb203518f0e771ef70c3610043e79d5524c06755bd5fe4d7e65c8eef
fixtures/empty_then_populated/scenario.json
float64_to_int64_fractional
float64 to int64 with fractional values
numeric
A float64 column becomes int64. v0 holds fractional values; each v1 value is the v0 value truncated toward zero. The values distinguish truncation from rounding: 1.5 -> 1, 2.999 -> 2, -1.5 -> -1.
[ "Truncation is what this producer did. It is not a statement of how a cast ought to behave; rounding, rejection and nulling are all observed in practice.", "NaN and infinities are deliberately absent: they have no int64 value, and including them would mix a second change into this scenario." ]
[ "change-type", "floating-point", "integer", "narrowing", "precision-loss" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 7, 7 ]
1
[ "change_type" ]
[ "amount" ]
[ "truncated_toward_zero" ]
true
true
true
fixtures/float64_to_int64_fractional/v0.parquet
fixtures/float64_to_int64_fractional/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"amount","nullable":true,"type":"double"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"amount","nullable":true,"type":"int64"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"amount","relation":"truncated_toward_zero"}]},"mutations":[{"kind":"change_type","new_type":"int64","...
sha256:166fb5770e018ab4f698312d8933ddc3e37b5ff37ed5a9a4529ebfda3fc380ff
sha256:15ee369c451d021b8d3ef93df90df199f0ac31de38f407a92e1e54e70f13697e
fixtures/float64_to_int64_fractional/scenario.json
float64_to_int64_integral
float64 to int64, integral values only
numeric
A float64 column becomes int64. Every non-null v0 value is integral and within range, so every value is unchanged. v0 contains -0.0, which equals 0 as a number.
[]
[ "change-type", "floating-point", "integer", "narrowing" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 7, 7 ]
1
[ "change_type" ]
[ "amount" ]
[]
true
false
true
fixtures/float64_to_int64_integral/v0.parquet
fixtures/float64_to_int64_integral/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"amount","nullable":true,"type":"double"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"amount","nullable":true,"type":"int64"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"int64","old_type":"double","path":"amount"}],"to":"v1"}]
sha256:a43ff1cbf7ce56cf5187a8115795f262f1be1e1569d3dec88e0d81bbd44afb29
sha256:d43070493146acfed94ca61906e05e2822ce6ce3bcc585525ef383b2444a3528
fixtures/float64_to_int64_integral/scenario.json
int32_to_decimal
int32 to decimal with every value unchanged
numeric
An int32 column becomes decimal128(12, 2), the smallest precision at scale 2 that holds every int32. Values, including both int32 extremes, are numerically unchanged (7 and 7.00 are the same number).
[ "The type moves across numeric families, from a binary integer to a scaled decimal, so it is not a width change within one family like int32 to int64.", "At the Parquet level the physical type changes from INT32 to FIXED_LEN_BYTE_ARRAY." ]
[ "change-type", "integer", "decimal", "widening", "numeric-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "quantity" ]
[]
true
false
true
fixtures/int32_to_decimal/v0.parquet
fixtures/int32_to_decimal/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"quantity","nullable":true,"type":"int32"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"quantity","nullable":true,"type":"decimal128(12, 2)"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"decimal128(12, 2)","old_type":"int32","path":"quantity"}],"to...
sha256:f96a0bbd72e6740792655d8e4c12117ebf1eed30a4c1d7509713de59322fb06f
sha256:d4a15d122b73125afebe1ab2d80efe7df252a9bb07450c17a255d2826c46bca8
fixtures/int32_to_decimal/scenario.json
int64_to_float64
int64 to float64 with values beyond 2^53
numeric
An int64 column becomes float64. Each v1 value is the nearest float64 to the v0 value. 2^53 and -2^63 are exactly representable and unchanged; 2^53 + 1 and the int64 maximum are not, so those two values change.
[ "The v1 value for the int64 maximum is 2^63, which is outside the int64 range: casting it back to int64 overflows." ]
[ "change-type", "integer", "floating-point", "precision-loss", "numeric-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 7, 7 ]
1
[ "change_type" ]
[ "measurement" ]
[ "nearest_float64" ]
true
true
true
fixtures/int64_to_float64/v0.parquet
fixtures/int64_to_float64/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"measurement","nullable":true,"type":"int64"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"measurement","nullable":true,"type":"double"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"measurement","relation":"nearest_float64"}]},"mutations":[{"kind":"change_type","new_type":"double","...
sha256:86f0e7ca2a815f9a6c7111edd0e533087dd00f2350da1d27a374c3f8e0e4aad0
sha256:07854510d0a27be81850e37bb14a963855c9cf2c7aa4faa08313bd11bf33375d
fixtures/int64_to_float64/scenario.json
list_element_widen
List elements widen from int32 to int64
structural
The element type of a list column changes from int32 to int64. Rows include a null list, an empty list and a list holding a null element, which are three different values. Every element value, including both int32 extremes, is unchanged.
[ "The element field name is not part of the corpus type: this writer stores a list written as list<item: ...> under the Parquet name 'element', and reads it back that way." ]
[ "list", "nested", "change-type", "integer", "widening", "nested-nulls" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "change_type" ]
[ "scores[]" ]
[]
true
false
true
fixtures/list_element_widen/v0.parquet
fixtures/list_element_widen/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"scores","nullable":true,"type":"list<int32>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"scores","nullable":true,"type":"list<int64>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"int64","old_type":"int32","path":"scores[]"}],"to":"v1"}]
sha256:b65cc8b84c69eed5f76eb35249b02c418df4cdfdff1d3b284cf3a0a0f0183715
sha256:509adc4dbda4efbb933f47fdb7ad6301b9ccf0990cfc338bc8921803fbb8fb81
fixtures/list_element_widen/scenario.json
map_value_widen
Map values widen from int32 to int64
structural
The value type of a map<string, int32> column becomes int64. Rows include a null map, an empty map, an empty-string key and a null value. Keys, values and entry order are unchanged.
[ "Map paths treat a map as a list of key/value entries, which is how Arrow and Parquet store it: the changed path is attributes[].value." ]
[ "map", "nested", "change-type", "integer", "widening" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "change_type" ]
[ "attributes[].value" ]
[]
true
false
true
fixtures/map_value_widen/v0.parquet
fixtures/map_value_widen/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"attributes","nullable":true,"type":"map<string, int32>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"attributes","nullable":true,"type":"map<string, int64>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"int64","old_type":"int32","path":"attributes[].value"}],"to":...
sha256:a4038108aaff3ba58f20879328c2eea4c3fc61ae1bda122ac33457e4f3c58343
sha256:67035ce47c8b7d7b13b868f9578a6c395bec92890ded44aa1d9153c58248cdba
fixtures/map_value_widen/scenario.json
nested_field_type_change
Nested fields widen from float32 to float64
structural
Both non-nullable fields of a nullable position struct change from float32 to float64. Every float32 value is exactly representable as float64 and is unchanged, which means the v1 value of 0.1 is 0.10000000149011612, not the float64 nearest to 0.1.
[ "One row has a null struct whose children are declared non-nullable: the children have no value in that row because their parent is null." ]
[ "struct", "nested", "change-type", "floating-point", "widening", "nested-nulls" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
2
[ "change_type" ]
[ "position.lat", "position.lon" ]
[]
true
false
true
fixtures/nested_field_type_change/v0.parquet
fixtures/nested_field_type_change/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"position","nullable":true,"type":"struct<lat: float not null, lon: float not null>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"position","nullable":true,"type":"struct<lat: double not null, lon: double not null>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"double","old_type":"float","path":"position.lat"},{"kind":"ch...
sha256:dcc1dfd854bf8926e56ca83cbd26ffe20d144fd6f98b468142a510c8f116b2ec
sha256:e189ef799a61251126841ba16c3433a2da8673c0f4d37b546b0bafa62c294967
fixtures/nested_field_type_change/scenario.json
null_type_to_string
Null-typed column becomes a string column
nullability
In v0 a column has the Arrow null type: it carries no values and no value type, which is what type inference produces from a batch where the column is always null. In v1 the same column is a string column with values.
[ "Parquet has no null type. The writer stores it as an optional INT32 column with the Null logical annotation, so the Parquet physical type changes as well (INT32 -> BYTE_ARRAY)." ]
[ "change-type", "null-type", "type-inference", "all-null" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "change_type" ]
[ "comment" ]
[ "nulls_filled" ]
true
true
true
fixtures/null_type_to_string/v0.parquet
fixtures/null_type_to_string/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"comment","nullable":true,"type":"null"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"comment","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"comment","relation":"nulls_filled"}]},"mutations":[{"kind":"change_type","new_type":"string","old_typ...
sha256:f4a5b5c1c1abb48a519b69e78cb14b2f17640c37877d9a724085108e7d2108e6
sha256:a73a821a2a5be6d2a0a270b14ebd3099a31cda3ea50df2c8961171d43cbf1e40
fixtures/null_type_to_string/scenario.json
nullable_column_backfilled
Nullable column loses its nulls; schema stays nullable
nullability
Every null in a nullable string column is filled in v1. The schema is identical in both versions: the field is still nullable although v1 contains no nulls. The empty string in v0 is a value, not a null, and is unchanged.
[ "Observing zero nulls in a batch says nothing about the declared nullability. Both files declare the column optional at the Arrow and Parquet levels." ]
[ "nullable", "backfill", "no-schema-change", "empty-string" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
0
[]
[ "email" ]
[ "nulls_filled" ]
false
true
false
fixtures/nullable_column_backfilled/v0.parquet
fixtures/nullable_column_backfilled/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"email","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"email","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"email","relation":"nulls_filled"}]},"mutations":[],"to":"v1"}]
sha256:7395491df667747e7bc840a44a8a9f7dd8341bde2e6625a32da0c6a7edd3ece8
sha256:04e5ebdbdf9144bc4866eaf5926038a0fc423b8e21d2c6f3e31bc6c4c1cf4b82
fixtures/nullable_column_backfilled/scenario.json
nullable_to_required
Nullable column becomes non-nullable
nullability
A nullable string column that happens to contain no nulls is declared non-nullable in v1. The data is identical; only the declaration changes (Parquet repetition optional -> required).
[]
[ "change-nullability", "string", "no-value-change" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_nullability" ]
[ "sku" ]
[]
true
false
true
fixtures/nullable_to_required/v0.parquet
fixtures/nullable_to_required/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"sku","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"sku","nullable":false,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_nullability","new_nullable":false,"old_nullable":true,"path":"sku"}],"to":"v1"}...
sha256:96301221808db936bc4757779a58856de486759ecc539da7c2763b6ffb1ff9cc
sha256:70f5dd82fa712a344d629f8e1333db830ebf3b3092eb9c405c9b26ccff09ec1c
fixtures/nullable_to_required/scenario.json
remove_nested_field
Remove a field from inside a struct
structural
The postcode field is removed from the address struct. Rows include a null struct, a struct whose street is null, and a struct of empty strings. Remaining nested values are unchanged.
[]
[ "struct", "nested", "remove-field", "nested-nulls" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "remove_field" ]
[ "address.postcode" ]
[]
true
false
true
fixtures/remove_nested_field/v0.parquet
fixtures/remove_nested_field/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"address","nullable":true,"type":"struct<street: string, city: string, postcode: string>"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"address","nullable":true,"type":"struct<street: string, city: string>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":false,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"remove_field","nullable":true,"path":"address.postcode","type":"string"}],"to":"v1"}]
sha256:ffc0803e00bf07b7da8ccb4f60df1121b9f160cf58c79b6a505b5450a2aeef28
sha256:fe2fd9399e594ce5ae78b4ba68f7dac7d891cef4ef9ab4eafc2742d1b7eb6cfc
fixtures/remove_nested_field/scenario.json
remove_nullable_column
Remove nullable column
subtractive
A nullable string column that held nulls, an empty string and values is removed. The remaining columns and all rows are unchanged.
[]
[ "remove-field", "string", "nullable" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "remove_field" ]
[ "middle_name" ]
[]
true
false
true
fixtures/remove_nullable_column/v0.parquet
fixtures/remove_nullable_column/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"middle_name","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":false,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"remove_field","nullable":true,"path":"middle_name","type":"string"}],"to":"v1"}]
sha256:f90aa90d7b9cd185f9473d9a41283a10cab10e82dc9519f22f3c9633de7d28d6
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
fixtures/remove_nullable_column/scenario.json
remove_required_column
Remove non-nullable populated column
subtractive
A non-nullable string column with a value in every row is removed. The remaining columns and all rows are unchanged.
[]
[ "remove-field", "string", "required" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "remove_field" ]
[ "country_code" ]
[]
true
false
true
fixtures/remove_required_column/v0.parquet
fixtures/remove_required_column/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"},{"name":"country_code","nullable":false,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"name","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":false,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"remove_field","nullable":false,"path":"country_code","type":"string"}],"to":"v1"}]
sha256:e9c57949b192096eb9a46534c80fabe05061656018bccc1fc10bb57ae82616bb
sha256:ffbd13144dd0c4b932b54b17badd51ef9e9209a5e577a165cab231d0fed9b6d2
fixtures/remove_required_column/scenario.json
rename_column_with_field_id
Rename column, identified by Parquet field id
structural
Column full_name is renamed to display_name. Both files carry Parquet field ids and the renamed column keeps field id 2, which is the evidence that this is a rename. Values are unchanged; they include 'Jose' with a composed and with a decomposed accent, which are different values.
[ "Without field ids a rename is physically indistinguishable from removing one column and adding another with the same values. The corpus only records a rename when the file carries evidence for it." ]
[ "rename-field", "field-id", "unicode" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "rename_field" ]
[ "display_name", "full_name" ]
[]
true
false
true
fixtures/rename_column_with_field_id/v0.parquet
fixtures/rename_column_with_field_id/v1.parquet
[{"field_id":1,"name":"id","nullable":false,"type":"int64"},{"field_id":2,"name":"full_name","nullable":true,"type":"string"},{"field_id":3,"name":"email","nullable":true,"type":"string"}]
[{"field_id":1,"name":"id","nullable":false,"type":"int64"},{"field_id":2,"name":"display_name","nullable":true,"type":"string"},{"field_id":3,"name":"email","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":false,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"field_id":2,"kind":"rename_field","old_path":"full_name","path":"display_name"}],"to":"v1"}]
sha256:bcc8c501e951e06ca6fca8d17fbb50f419f9bccb5c65d5b3a13b6b6b9fc1e99c
sha256:7da200928b89a496a2b3147837f7860b8c400974d0a7a4af56084d64a008ef4e
fixtures/rename_column_with_field_id/scenario.json
reorder_columns
Reorder columns
structural
The same four columns appear in a different order. Names, types, nullability and values are unchanged. Neither order is alphabetical, so a reader that sorts columns by name disagrees with both.
[]
[ "reorder-fields", "column-order" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "reorder_fields" ]
[]
[]
true
false
true
fixtures/reorder_columns/v0.parquet
fixtures/reorder_columns/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"region","nullable":true,"type":"string"},{"name":"amount","nullable":true,"type":"int64"},{"name":"active","nullable":true,"type":"bool"}]
[{"name":"amount","nullable":true,"type":"int64"},{"name":"id","nullable":false,"type":"int64"},{"name":"active","nullable":true,"type":"bool"},{"name":"region","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"reorder_fields","new_order":["amount","id","active","region"],"old_order":["id","regio...
sha256:e7bd349f7eae6691f2bc0ab2a70ac040ca666f58e8acd691ed32436048545f03
sha256:a514e4bbb4aee2c15f1df67a2100def2bedf8bd85282eafede8c60b4f1ca2c0c
fixtures/reorder_columns/scenario.json
required_to_nullable
Non-nullable column becomes nullable
nullability
A non-nullable string column is declared nullable in v1. The data is identical and contains no nulls in either version; only the declaration changes (Parquet repetition required -> optional).
[]
[ "change-nullability", "string", "no-value-change" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_nullability" ]
[ "sku" ]
[]
true
false
true
fixtures/required_to_nullable/v0.parquet
fixtures/required_to_nullable/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"sku","nullable":false,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"sku","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_nullability","new_nullable":true,"old_nullable":false,"path":"sku"}],"to":"v1"}...
sha256:70f5dd82fa712a344d629f8e1333db830ebf3b3092eb9c405c9b26ccff09ec1c
sha256:96301221808db936bc4757779a58856de486759ecc539da7c2763b6ffb1ff9cc
fixtures/required_to_nullable/scenario.json
rows_appended_without_key
Rows appended to a dataset without a key
edge
An append-only log with no identifying column gains two rows. The first four rows, which include two identical rows, are unchanged and in the same order. The schema is unchanged.
[ "row_identity is empty: rows are identified by position, so row i of v0 is row i of v1. The duplicate rows are why no column set can serve as a key." ]
[ "keyless", "append", "duplicate-rows", "no-schema-change" ]
[]
2
[ "v0", "v1" ]
[ 4, 6 ]
0
[]
[]
[]
false
false
false
fixtures/rows_appended_without_key/v0.parquet
fixtures/rows_appended_without_key/v1.parquet
[{"name":"level","nullable":true,"type":"string"},{"name":"message","nullable":true,"type":"string"}]
[{"name":"level","nullable":true,"type":"string"},{"name":"message","nullable":true,"type":"string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":false,"row_identity_preserved":false,"rows_retained":true,"value_changes":[]},"mutations":[],"to":"v1"}]
sha256:87f3ef6e927225e236d82b394450de95324e6d095adba8279801412fe80c2961
sha256:85f671701106ace08b2863716021d6bc28df921f1fbf19b6c3d243b88034a290
fixtures/rows_appended_without_key/scenario.json
rows_deleted
Rows deleted from a keyed dataset
edge
Two of six rows, ids 2 and 5, are absent from v1. The four remaining rows keep their values and their relative order. The schema is unchanged.
[ "rows_retained is false. common_values_preserved quantifies over the rows present in both versions, and those rows are unchanged.", "v1 has no rows for ids 2 and 5, rather than rows with nulls or empty values. The deleted rows held an empty status (id 2) and a null amount (id 5)." ]
[ "delete", "keyed", "no-schema-change" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 4 ]
0
[]
[]
[]
false
false
false
fixtures/rows_deleted/v0.parquet
fixtures/rows_deleted/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"string"},{"name":"amount","nullable":true,"type":"int64"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"string"},{"name":"amount","nullable":true,"type":"int64"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":false,"row_identity_preserved":false,"rows_retained":false,"value_changes":[]},"mutations":[],"to":"v1"}]
sha256:190cc5f8270f318300bfce01b7d28a81df23829c481f97057fdfe866d050e976
sha256:9d4eb19479eb6a32bf96d5800da5cf576a00a58dfdbeedb920a7ed9bf8106336
fixtures/rows_deleted/scenario.json
string_to_binary
string to binary holding the UTF-8 bytes
representation
A string column becomes a binary column. Each v1 value is the UTF-8 encoding of the v0 value. Text and bytes are different kinds of value, so values are not preserved even though the stored bytes are identical. The composed and decomposed 'cafe' encode to different bytes.
[ "At the Parquet level both columns are BYTE_ARRAY; only the String logical annotation is removed." ]
[ "change-type", "string", "binary", "unicode" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "payload" ]
[ "utf8_encoded" ]
true
true
true
fixtures/string_to_binary/v0.parquet
fixtures/string_to_binary/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"payload","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"payload","nullable":true,"type":"binary"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"payload","relation":"utf8_encoded"}]},"mutations":[{"kind":"change_type","new_type":"binary","old_typ...
sha256:1d2caddd132217e781f4c0af6de1abd1d098bb9c58e2716ea2b8041d758d6c56
sha256:04a2d7a0380eb970df52303bec4f6e581077e70401e730f8926d84f2e4fa0919
fixtures/string_to_binary/scenario.json
string_to_dictionary
string to dictionary-encoded string
representation
A plain string column becomes a dictionary<int8, string> column, as written for a pandas categorical. v0 is written without dictionary encoding; v1 with it. Decoded values are unchanged.
[ "The Arrow type change is recorded only in the stored Arrow schema. The Parquet schema is identical; the Parquet difference is in the column chunk encodings (PLAIN vs RLE_DICTIONARY).", "Readers that ignore the stored Arrow schema may decode v1 as plain strings, or as a dictionary with int32 indices. The int8 ind...
[ "change-type", "dictionary", "categorical", "encoding", "arrow-only" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "status" ]
[]
true
false
false
fixtures/string_to_dictionary/v0.parquet
fixtures/string_to_dictionary/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"status","nullable":true,"type":"dictionary<values=string, indices=int8, ordered=0>"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"dictionary<values=string, indices=int8, ordered=0>","old_type"...
sha256:e5b66495d62b534fd0a382904d46f85c23aaf18721089e242e90676fdb4d9444
sha256:6dfdbf0a9de0eac2a01b3d7a88cb06abe18009dfa4d033259c14f39d7bf16d9b
fixtures/string_to_dictionary/scenario.json
string_to_large_string
string to large_string
representation
A string column becomes large_string (64-bit offsets). Values are unchanged; they include an empty string, whitespace, the literal text NULL, CJK, an emoji, and 'cafe' with the accent both composed (U+00E9) and decomposed (e + U+0301).
[ "The difference exists only in the stored Arrow schema. Both files have the same Parquet schema (BYTE_ARRAY, String), so a reader that ignores the Arrow schema sees no change at all." ]
[ "change-type", "string", "offsets", "unicode", "arrow-only" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 8, 8 ]
1
[ "change_type" ]
[ "label" ]
[]
true
false
false
fixtures/string_to_large_string/v0.parquet
fixtures/string_to_large_string/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"label","nullable":true,"type":"string"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"label","nullable":true,"type":"large_string"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":true,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"large_string","old_type":"string","path":"label"}],"to":"v1"}]
sha256:af037b683e64ba1bd9e63dd778b55bea85f8e898f85becfa6f4450ba01a036c7
sha256:4ed55bd1dd61a05cd056a56e3d0554a6f764bd6a03522e3189372a421d692f9d
fixtures/string_to_large_string/scenario.json
timestamp_naive_to_utc
Naive timestamp becomes UTC timestamp
temporal
A timestamp column without a time zone becomes timestamp with tz=UTC. The stored integers are identical, so each wall-clock value in v0 is reinterpreted as the same wall-clock value in UTC. A naive value and a zoned value are not the same value.
[ "At the Parquet level this is the isAdjustedToUTC flag of the Timestamp logical type changing from false to true; the physical INT64 values are unchanged.", "Values include one microsecond before the epoch, the epoch, a leap day and a wall-clock time that does not exist in Central European time." ]
[ "change-type", "timestamp", "timezone", "utc" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 5, 5 ]
1
[ "change_type" ]
[ "event_time" ]
[ "interpreted_as_utc" ]
true
true
true
fixtures/timestamp_naive_to_utc/v0.parquet
fixtures/timestamp_naive_to_utc/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"event_time","nullable":true,"type":"timestamp[us]"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"event_time","nullable":true,"type":"timestamp[us, tz=UTC]"}]
[{"from":"v0","invariants":{"common_values_preserved":false,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[{"path":"event_time","relation":"interpreted_as_utc"}]},"mutations":[{"kind":"change_type","new_type":"timesta...
sha256:ad6f3633367e76d56dca96b7ca6530e97c855cf26d71ccb0264a35da1752fc58
sha256:fa7de3993fb420dfd70186ace21f0493654625a6c25682e648df34380844011b
fixtures/timestamp_naive_to_utc/scenario.json
timestamp_ns_to_us
Timestamp unit changes from nanoseconds to microseconds
temporal
A naive timestamp column changes unit from nanoseconds to microseconds. Every v0 value is a whole number of microseconds, so every instant is unchanged. Values include both ends of the range timestamp[ns] can hold, one microsecond before the epoch, and a leap day.
[ "At the Parquet level the Timestamp logical type's unit changes from NANOS to MICROS and every stored INT64 is divided by 1000.", "Timestamps compare by instant, independent of unit, so no value changes. The v0 extremes sit at the edges of the int64 nanosecond range; the microsecond range is 1000 times wider." ]
[ "change-type", "timestamp", "unit", "nanoseconds", "range-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "event_time" ]
[]
true
false
true
fixtures/timestamp_ns_to_us/v0.parquet
fixtures/timestamp_ns_to_us/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"event_time","nullable":true,"type":"timestamp[ns]"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"event_time","nullable":true,"type":"timestamp[us]"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"timestamp[us]","old_type":"timestamp[ns]","path":"event_time"...
sha256:fb6249b26547f63e125784d224fe1f57cd43eb540150f0a60fa28f39d7ba3cd9
sha256:78a16095bb6fb126b9cd3f13f993c4d189659a7b963c08d708ad3586ea9b313a
fixtures/timestamp_ns_to_us/scenario.json
widen_int32_to_int64
Widen int32 to int64
numeric
An int32 column becomes int64. Values, including both int32 extremes, zero and a null, are unchanged.
[]
[ "change-type", "integer", "widening", "numeric-extremes" ]
[ "id" ]
2
[ "v0", "v1" ]
[ 6, 6 ]
1
[ "change_type" ]
[ "quantity" ]
[]
true
false
true
fixtures/widen_int32_to_int64/v0.parquet
fixtures/widen_int32_to_int64/v1.parquet
[{"name":"id","nullable":false,"type":"int64"},{"name":"quantity","nullable":true,"type":"int32"}]
[{"name":"id","nullable":false,"type":"int64"},{"name":"quantity","nullable":true,"type":"int64"}]
[{"from":"v0","invariants":{"common_values_preserved":true,"parquet_schema_preserved":false,"paths_retained":true,"row_count_preserved":true,"row_identity_preserved":true,"rows_retained":true,"value_changes":[]},"mutations":[{"kind":"change_type","new_type":"int64","old_type":"int32","path":"quantity"}],"to":"v1"}]
sha256:ac3c16abadb0598042a7d24873373bcd88ed27198890526a07e71caa6469ba4e
sha256:49ad7406c3512c129d76d7a58b9c29d2298363952dab9cfb529fd61777c23db7
fixtures/widen_int32_to_int64/scenario.json

Tabular Evolution Corpus

Source, generator and validator: github.com/AndreaBozzo/tabular-evolution-corpus. This card is the README.md of the Hugging Face dataset repository; the GitHub repository keeps its own README.

A small, deterministic corpus of Parquet datasets that change over time. Each scenario is one logical dataset at two or more versions, with a manifest that records what structurally changed and which row and value facts hold across each transition. Every fact is verified against the files.

The corpus records facts, not compatibility verdicts. Whether a change is acceptable is for the engine, table format or data contract under test to decide.

What the viewer shows

The viewer shows the catalog, data/scenarios.parquet, with one row per scenario (35 rows). The catalog is an index. The versioned fixture files it points to live in the same repository under fixtures/<scenario_id>/ (v0.parquet, v1.parquet, ..., scenario.json).

The configs block above deliberately restricts the viewer and load_dataset to the catalog. The fixture Parquet files have a different schema in every scenario and are not rows of one dataset. Without the explicit config, automatic data-file detection would try to load them as one.

Catalog columns

column type meaning
scenario_id string stable identifier; also the fixture directory name
title, description string what changes, stated as facts
category string additive, subtractive, numeric, nullability, representation, structural, temporal, edge
notes list caveats, e.g. differences that exist only in the stored Arrow schema
tags list free-form, kebab-case
row_identity list columns that identify a row in every version; empty means rows are identified by position
version_count, version_ids, row_counts int64, list, list versions in order
mutation_count, mutation_kinds int64, list structural changes across all transitions
paths list field paths touched by a mutation or a value change
value_change_relations list how changed values relate (nearest_float64, utf8_encoded, ...)
schema_changed, values_changed, parquet_schema_changed bool quick filters
base_file, final_file, manifest_path string repository-relative paths
base_schema_json, final_schema_json, transitions_json string JSON; full detail is in scenario.json
base_fingerprint, final_fingerprint string logical fingerprints (schema + values + row order)

Nested manifest structures are JSON strings because their shape depends on the mutation kind. Every other column is a scalar or a list of scalars.

Usage

from datasets import load_dataset

catalog = load_dataset("AndreaBozzo/tabular-evolution-corpus", split="scenarios")

To get the fixtures, download the repository files:

from huggingface_hub import snapshot_download

root = snapshot_download("AndreaBozzo/tabular-evolution-corpus", repo_type="dataset")
# root/fixtures/<scenario_id>/{v0.parquet, v1.parquet, scenario.json}

Every corpus release is a tag, both here and on GitHub. Pass revision="v0.2.0" to either call to get exactly that release. A released tag is never moved, and a changed corpus gets a new release.

-- DuckDB
SELECT scenario_id, mutation_kinds, row_counts
FROM read_parquet('hf://datasets/AndreaBozzo/tabular-evolution-corpus/data/scenarios.parquet')
WHERE category = 'numeric';

Dataset structure

data/scenarios.parquet          catalog (this viewer)
fixtures/<scenario_id>/         one directory per scenario
    v0.parquet, v1.parquet ...  versions, 0-8 rows each
    scenario.json               manifest (JSON Schema: schema/scenario.schema.json)
schema/scenario.schema.json
SHA256SUMS

Creation

All data is synthetic, written by the generator in the source repository with a pinned PyArrow Parquet writer (uncompressed, format 2.6, one row group). Values are chosen to expose boundaries: NaN, infinities, -0.0, 2^53 + 1, int and decimal extremes, composed vs decomposed Unicode, empty strings, nested nulls, pre-epoch timestamps, timestamps at both ends of the nanosecond range. Regeneration is logically reproducible, and byte-reproducible with the same PyArrow version.

Intended use and limitations

Intended as test input for ingestion frameworks, dataframe libraries, query engines, lakehouse tooling, schema registries, contract validators and profilers. It is not training data, not a performance benchmark, and not a conformance suite for Arrow or Parquet (see apache/parquet-testing and the Arrow integration tests for those). Corpus 0.2.0 covers Parquet only and 35 scenarios. Nested evolution covers structs, lists and maps; datasets are keyed or identified by row position.

One writer. Every file is written by a single pinned Parquet writer (PyArrow / Arrow C++, recorded in each file's created_by). The scenarios vary the data, not the writer. Differences that exist only in the Arrow schema stored in the file (large_string, dictionary encoding) are flagged in each manifest, because readers that ignore that schema see no change. Fixtures from other writers are planned but not included.

License

Apache-2.0. No personal data: every name, address and e-mail in the fixtures is invented or a well-known public name used as a placeholder.

Downloads last month
87