Skip to content

Support for 'time' type in Iceberg, read and write #1731

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ianton-ru wants to merge 8 commits into
base: antalya-26.1
Choose a base branch
from
Open

Support for 'time' type in Iceberg, read and write #1731

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
9f00f37
Iceberg type 'time' support
ianton-ru Mar 18, 2026
bb896c3
Fix tests
ianton-ru Mar 19, 2026
28ba40b
Fix
ianton-ru Mar 19, 2026
f0e1038
Fix test
ianton-ru Apr 2, 2026
c2b8ac0
Write time to Iceberg
ianton-ru Apr 8, 2026
ced1375
Proper write metadata to Avro
ianton-ru Apr 10, 2026
40b252b
Small fixes
ianton-ru May 5, 2026
0ef39ae
Fixes, AI assist (Cursor + ChatGPT 5.5)
ianton-ru May 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 13 additions & 1 deletion src/Processors/Formats/Impl/AvroRowInputFormat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,6 +156,9 @@ static void insertNumber(IColumn & column, WhichDataType type, T value)
case TypeIndex::DateTime64:
assert_cast<ColumnDecimal<DateTime64> &>(column).insertValue(static_cast<Int64>(value));
break;
case TypeIndex::Time64:
assert_cast<ColumnDecimal<Time64> &>(column).insertValue(static_cast<Int64>(value));
break;
case TypeIndex::IPv4:
assert_cast<ColumnIPv4 &>(column).insertValue(IPv4(static_cast<UInt32>(value)));
break;
Expand Down Expand Up @@ -304,6 +307,8 @@ AvroDeserializer::DeserializeFn AvroDeserializer::createDeserializeFn(const avro
return createDecimalDeserializeFn<DataTypeDecimal256>(root_node, target_type, false);
if (target.isDateTime64())
return createDecimalDeserializeFn<DataTypeDateTime64>(root_node, target_type, false);
if (target.isTime64())
return createDecimalDeserializeFn<DataTypeTime64>(root_node, target_type, false);
break;
case avro::AVRO_INT:
if (target_type->isValueRepresentedByNumber())
Expand Down Expand Up @@ -1283,8 +1288,11 @@ DataTypePtr AvroSchemaReader::avroNodeToDataType(avro::NodePtr node)
{
case avro::Type::AVRO_INT:
{
if (node->logicalType().type() == avro::LogicalType::DATE)
auto logical_type = node->logicalType();
if (logical_type.type() == avro::LogicalType::DATE)
return {std::make_shared<DataTypeDate32>()};
if (logical_type.type() == avro::LogicalType::TIME_MILLIS)
return {std::make_shared<DataTypeTime64>(3)};

return {std::make_shared<DataTypeInt32>()};
}
Expand All @@ -1295,6 +1303,10 @@ DataTypePtr AvroSchemaReader::avroNodeToDataType(avro::NodePtr node)
return {std::make_shared<DataTypeDateTime64>(3)};
if (logical_type.type() == avro::LogicalType::TIMESTAMP_MICROS)
return {std::make_shared<DataTypeDateTime64>(6)};
if (logical_type.type() == avro::LogicalType::TIME_MILLIS)
return {std::make_shared<DataTypeTime64>(3)};
if (logical_type.type() == avro::LogicalType::TIME_MICROS)
return {std::make_shared<DataTypeTime64>(6)};

return std::make_shared<DataTypeInt64>();
}
Expand Down
50 changes: 50 additions & 0 deletions src/Processors/Formats/Impl/Parquet/PrepareForWrite.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@
#include <DataTypes/DataTypeMap.h>
#include <DataTypes/DataTypeDateTime64.h>
#include <DataTypes/DataTypeFixedString.h>
#include <DataTypes/DataTypeTime64.h>
#include <DataTypes/DataTypeCustom.h>

/// This file deals with schema conversion and with repetition and definition levels.
Expand Down Expand Up @@ -475,6 +476,55 @@ void preparePrimitiveColumn(ColumnPtr column, DataTypePtr type, const std::strin
case TypeIndex::Decimal128: decimal(16, getDecimalPrecision(*type), getDecimalScale(*type)); break;
case TypeIndex::Decimal256: decimal(32, getDecimalPrecision(*type), getDecimalScale(*type)); break;

case TypeIndex::Time:
{
parq::TimeUnit unit;
unit.__set_MICROS({});

parq::TimeType tt;
tt.__set_isAdjustedToUTC(false);
tt.__set_unit(unit);

parq::LogicalType t;
t.__set_TIME(tt);
types(T::INT64, parq::ConvertedType::TIME_MICROS, t);
state.datetime_multiplier = 1'000'000;
break;
}

case TypeIndex::Time64:
{
std::optional<parq::ConvertedType::type> converted;
parq::TimeUnit unit;
const auto & dt = assert_cast<const DataTypeTime64 &>(*type);
UInt32 scale = dt.getScale();
UInt32 converted_scale;
if (scale <= 6)
{
converted = parq::ConvertedType::TIME_MICROS;
unit.__set_MICROS({});
converted_scale = 6;
}
else if (scale <= 9)
{
unit.__set_NANOS({});
converted_scale = 9;
}
else
{
throw Exception(ErrorCodes::LOGICAL_ERROR, "Unexpected Time64 scale: {}", scale);
}

parq::TimeType tt;
tt.__set_isAdjustedToUTC(false);
tt.__set_unit(unit);
parq::LogicalType t;
t.__set_TIME(tt);
types(T::INT64, converted, t);
state.datetime_multiplier = DataTypeTime64::getScaleMultiplier(converted_scale - scale);
break;
}

default:
throw Exception(ErrorCodes::UNKNOWN_TYPE, "Internal type '{}' of column '{}' is not supported for conversion into Parquet data format.", type->getFamilyName(), name);
}
Expand Down
51 changes: 45 additions & 6 deletions src/Processors/Formats/Impl/Parquet/Write.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -358,28 +358,32 @@ struct ConverterNumeric
}
};

struct ConverterDateTime64WithMultiplier
template <typename TYPE>
struct ConverterTimeType64WithMultiplierImpl
{
using Statistics = StatisticsNumeric<Int64, Int64>;

using Col = ColumnDecimal<DateTime64>;
using Col = ColumnDecimal<TYPE>;
const Col & column;
Int64 multiplier;
PODArray<Int64> buf;

ConverterDateTime64WithMultiplier(const ColumnPtr & c, Int64 multiplier_) : column(assert_cast<const Col &>(*c)), multiplier(multiplier_) {}
ConverterTimeType64WithMultiplierImpl(const ColumnPtr & c, Int64 multiplier_) : column(assert_cast<const Col &>(*c)), multiplier(multiplier_) {}

const Int64 * getBatch(size_t offset, size_t count)
{
buf.resize(count);
for (size_t i = 0; i < count; ++i)
/// Not checking overflow because DateTime64 values should already be in the range where
/// Not checking overflow because Time64/DateTime64 values should already be in the range where
/// they fit in Int64 at any allowed scale (i.e. up to nanoseconds).
buf[i] = column.getData()[offset + i].value * multiplier;
return buf.data();
}
};

using ConverterDateTime64WithMultiplier = ConverterTimeType64WithMultiplierImpl<DateTime64>;
using ConverterTime64WithMultiplier = ConverterTimeType64WithMultiplierImpl<Time64>;

/// Multiply DateTime by 1000 to get milliseconds (because Parquet doesn't support seconds).
struct ConverterDateTime
{
Expand All @@ -400,6 +404,26 @@ struct ConverterDateTime
}
};

struct ConverterTime
{
using Statistics = StatisticsNumeric<Int64, Int64>;

using Col = ColumnVector<Int32>;
const Col & column;
PODArray<Int64> buf;
Int64 multiplier;

ConverterTime(const ColumnPtr & c, Int64 multiplier_) : column(assert_cast<const Col &>(*c)), multiplier(multiplier_) {}

const Int64 * getBatch(size_t offset, size_t count)
{
buf.resize(count);
for (size_t i = 0; i < count; ++i)
buf[i] = static_cast<Int64>(column.getData()[offset + i]) * multiplier;
return buf.data();
}
};

struct ConverterString
{
using Statistics = StatisticsStringRef;
Expand Down Expand Up @@ -1164,7 +1188,12 @@ void writeColumnChunkBody(
N(Int16, Int32Type);
break;
}
case TypeIndex::Int32 : N(Int32, Int32Type); break;
case TypeIndex::Int32:
if (s.type->getTypeId() == TypeIndex::Time)
writeColumnImpl<parquet::Int64Type>(s, options, out, ConverterTime(s.primitive_column, s.datetime_multiplier));
else
N(Int32, Int32Type);
break;
case TypeIndex::Int64 : N(Int64, Int64Type); break;

case TypeIndex::UInt32:
Expand All @@ -1177,7 +1206,6 @@ void writeColumnChunkBody(
writeColumnImpl<parquet::Int64Type>(s, options, out, ConverterDateTime(s.primitive_column));
}
break;

#undef N

case TypeIndex::Float32:
Expand Down Expand Up @@ -1246,6 +1274,17 @@ void writeColumnChunkBody(
case TypeIndex::Decimal256: D(Decimal256); break;
#undef D

case TypeIndex::Time64:
if (s.datetime_multiplier == 1)
writeColumnImpl<parquet::Int64Type>(
s, options, out, ConverterNumeric<ColumnDecimal<Time64>, Int64, Int64>(
s.primitive_column));
else
writeColumnImpl<parquet::Int64Type>(
s, options, out, ConverterTime64WithMultiplier(
s.primitive_column, s.datetime_multiplier));
break;

default:
throw Exception(ErrorCodes::LOGICAL_ERROR, "Unexpected column type: {}", s.primitive_column->getFamilyName());
}
Expand Down
1 change: 1 addition & 0 deletions src/Storages/ObjectStorage/DataLakes/Iceberg/Constant.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@ DEFINE_ICEBERG_FIELD(time);
DEFINE_ICEBERG_FIELD(timestamp);
DEFINE_ICEBERG_FIELD(timestamptz);
DEFINE_ICEBERG_FIELD(type)
DEFINE_ICEBERG_FIELD(logicalType); /// this field has a camelCase name
DEFINE_ICEBERG_FIELD(transform);
DEFINE_ICEBERG_FIELD(direction);

Expand Down
67 changes: 65 additions & 2 deletions src/Storages/ObjectStorage/DataLakes/Iceberg/IcebergWrites.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@
#include <Core/TypeId.h>
#include <DataTypes/DataTypeNullable.h>
#include <DataTypes/DataTypeString.h>
#include <DataTypes/DataTypeTime64.h>
#include <DataTypes/DataTypesNumber.h>
#include <DataTypes/IDataType.h>
#include <Databases/DataLake/Common.h>
Expand Down Expand Up @@ -129,6 +130,8 @@ bool canDumpIcebergStats(const Field & field, DataTypePtr type)
case TypeIndex::Date32:
case TypeIndex::Int64:
case TypeIndex::DateTime64:
case TypeIndex::Time:
case TypeIndex::Time64:
case TypeIndex::String:
return true;
default:
Expand All @@ -144,6 +147,46 @@ std::vector<uint8_t> dumpValue(T value)
return bytes;
}

DataTypePtr getTimeTypeOrNull(DataTypePtr type)
{
if (type->isNullable())
return getTimeTypeOrNull(assert_cast<const DataTypeNullable *>(type.get())->getNestedType());

const WhichDataType which(type);
if (which.isTime() || which.isTime64())
return type;

return nullptr;
}

Int64 getTimeValueInMicroseconds(const Field & field, DataTypePtr type)
{
if (type->isNullable())
return getTimeValueInMicroseconds(field, assert_cast<const DataTypeNullable *>(type.get())->getNestedType());

const WhichDataType which(type);
if (which.isTime())
{
if (field.getType() == Field::Types::Int64)
return field.safeGet<Int64>() * 1'000'000;
if (field.getType() == Field::Types::UInt64)
return static_cast<Int64>(field.safeGet<UInt64>()) * 1'000'000;
return static_cast<Int64>(field.safeGet<Int32>()) * 1'000'000;
}

if (which.isTime64())
{
const auto scale = getDecimalScale(*type);
if (scale > 6)
throw Exception(ErrorCodes::BAD_ARGUMENTS, "Unsupported type for iceberg {}", type->getName());

const auto value = field.safeGet<Decimal64>().getValue().value;
return value * DataTypeTime64::getScaleMultiplier(6 - scale).value;
}

throw Exception(ErrorCodes::LOGICAL_ERROR, "Expected Time or Time64, got {}", type->getName());
}

std::vector<uint8_t> dumpFieldToBytes(const Field & field, DataTypePtr type)
{
switch (type->getTypeId())
Expand All @@ -154,8 +197,12 @@ std::vector<uint8_t> dumpFieldToBytes(const Field & field, DataTypePtr type)
case TypeIndex::Date:
case TypeIndex::Date32:
return dumpValue(field.safeGet<Int32>());
case TypeIndex::Time:
return dumpValue(getTimeValueInMicroseconds(field, type));
case TypeIndex::Int64:
return dumpValue(field.safeGet<Int64>());
case TypeIndex::Time64:
return dumpValue(getTimeValueInMicroseconds(field, type));
case TypeIndex::DateTime64:
return dumpValue(field.safeGet<Decimal64>().getValue().value);
case TypeIndex::String:
Expand Down Expand Up @@ -219,7 +266,16 @@ void extendSchemaForPartitions(
Poco::JSON::Object::Ptr field = new Poco::JSON::Object;
field->set(Iceberg::f_field_id, 1000 + i);
field->set(Iceberg::f_name, partition_columns[i]);
field->set(Iceberg::f_type, getAvroType(partition_types[i]));
auto logical_type = getAvroLogicalType(partition_types[i]);
if (!logical_type.isEmpty())
{
Poco::JSON::Object::Ptr type_field = new Poco::JSON::Object;
type_field->set(Iceberg::f_type, getAvroType(partition_types[i]));
type_field->set(Iceberg::f_logicalType, logical_type);
field->set(Iceberg::f_type, type_field);
}
else
field->set(Iceberg::f_type, getAvroType(partition_types[i]));
partition_fields->add(field);
}

Expand Down Expand Up @@ -385,6 +441,14 @@ void generateManifestFile(
avro::GenericRecord & partition_record = data_file.field("partition").value<avro::GenericRecord>();
for (size_t i = 0; i < partition_columns.size(); ++i)
{
auto partition_time_type = getTimeTypeOrNull(partition_types[i]);
if (!partition_values[i].isNull() && partition_time_type)
{
partition_record.field(partition_columns[i]) =
avro::GenericDatum(getTimeValueInMicroseconds(partition_values[i], partition_types[i]));
continue;
}

switch (partition_values[i].getType())
{
case Field::Types::Int64:
Expand Down Expand Up @@ -412,7 +476,6 @@ void generateManifestFile(
partition_record.field(partition_columns[i]) =
avro::GenericDatum(partition_values[i].safeGet<Decimal64>().getValue());
break;

case Field::Types::Null:
break;

Expand Down
2 changes: 1 addition & 1 deletion src/Storages/ObjectStorage/DataLakes/Iceberg/SchemaProcessor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,7 +243,7 @@ DataTypePtr IcebergSchemaProcessor::getSimpleType(const String & type_name, Cont
if (type_name == f_date)
return std::make_shared<DataTypeDate32>();
if (type_name == f_time)
return std::make_shared<DataTypeInt64>();
return std::make_shared<DataTypeTime64>(6);
Comment on lines 245 to +246
Copy link
Copy Markdown

@chatgpt-codex-connector chatgpt-codex-connector Bot May 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Decode Iceberg TIME without timezone-dependent conversion

Mapping Iceberg time to Time64(6) here now routes Parquet TIME columns through the existing Parquet decoder path that still infers DateTime64 for logical TIME (SchemaConverter.cpp around logical.__isset.TIME), and then relies on cast to Time64; that cast applies a timezone offset (FunctionsConversion.h computes local seconds-of-day), so values shift in non-UTC environments. In practice, a stored 12:00:00 TIME can be read as 07:00:00 on America/New_York, which breaks correctness for Iceberg TIME reads and partition filtering.

Useful? React with 👍 / 👎.

if (type_name == f_timestamp)
return std::make_shared<DataTypeDateTime64>(6);
if (type_name == f_timestamptz)
Expand Down
Loading
Loading