Small improvements in JSON schema inference

This commit is contained in:
avogar 2024-03-04 17:32:22 +00:00
parent fb02137fcc
commit 70abdf7a41
11 changed files with 142 additions and 41 deletions

View File

@ -549,6 +549,48 @@ Result:
└───────┴─────────────────────────────────────────────────────────────────────────────────────────────────┴──────────────┴────────────────────┴─────────┴──────────────────┴────────────────┘
```
##### input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects
Enabling this setting allows to use String type for ambiguous paths during named tuples inference from JSON objects (when `input_format_json_try_infer_named_tuples_from_objects` is enabled) instead of an exception.
It allows to read JSON objects as named Tuples even if there are ambiguous paths.
Disabled by default.
**Examples**
With disabled setting:
```sql
SET input_format_json_try_infer_named_tuples_from_objects = 1;
SET input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects = 0;
DESC format(JSONEachRow, '{"obj" : {"a" : 42}}, {"obj" : {"a" : {"b" : "Hello"}}}');
```
Result:
```text
Code: 636. DB::Exception: The table structure cannot be extracted from a JSONEachRow format file. Error:
Code: 117. DB::Exception: JSON objects have ambiguous paths: 'a' (with type Int64) and 'a.b'. You can enable setting input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects to use String type for path 'a'. (INCORRECT_DATA) (version 24.3.1.1).
You can specify the structure manually. (CANNOT_EXTRACT_TABLE_STRUCTURE)
```
With enabled setting:
```sql
SET input_format_json_try_infer_named_tuples_from_objects = 1;
SET input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects = 1;
DESC format(JSONEachRow, '{"obj" : "a" : 42}, {"obj" : {"a" : {"b" : "Hello"}}}');
SELECT * FROM format(JSONEachRow, '{"obj" : {"a" : 42}}, {"obj" : {"a" : {"b" : "Hello"}}}');
```
Result:
```text
┌─name─┬─type──────────────────────────┬─default_type─┬─default_expression─┬─comment─┬─codec_expression─┬─ttl_expression─┐
│ obj │ Tuple(a Nullable(String)) │ │ │ │ │ │
└──────┴───────────────────────────────┴──────────────┴────────────────────┴─────────┴──────────────────┴────────────────┘
┌─obj─────────────────┐
│ ('42') │
│ ('{"b" : "Hello"}') │
└─────────────────────┘
```
##### input_format_json_read_objects_as_strings
Enabling this setting allows reading nested JSON objects as strings.
@ -1554,6 +1596,28 @@ DESC format(JSONEachRow, $$
└──────┴──────────────────┴──────────────┴────────────────────┴─────────┴──────────────────┴────────────────┘
```
#### input_format_try_infer_exponent_floats
If enabled, ClickHouse will try to infer floats in exponential form for text formats (except JSON where numbers in exponential form are always inferred).
Disabled by default.
**Example**
```sql
SET input_format_try_infer_exponent_floats = 1;
DESC format(CSV,
$$1.1E10
2.3e-12
42E00
$$)
```
```response
┌─name─┬─type──────────────┬─default_type─┬─default_expression─┬─comment─┬─codec_expression─┬─ttl_expression─┐
│ c1 │ Nullable(Float64) │ │ │ │ │ │
└──────┴───────────────────┴──────────────┴────────────────────┴─────────┴──────────────────┴────────────────┘
```
## Self describing formats {#self-describing-formats}
Self-describing formats contain information about the structure of the data in the data itself,

View File

@ -1018,6 +1018,7 @@ class IColumn;
M(Bool, input_format_json_read_objects_as_strings, true, "Allow to parse JSON objects as strings in JSON input formats", 0) \
M(Bool, input_format_json_read_arrays_as_strings, true, "Allow to parse JSON arrays as strings in JSON input formats", 0) \
M(Bool, input_format_json_try_infer_named_tuples_from_objects, true, "Try to infer named tuples from JSON objects in JSON input formats", 0) \
M(Bool, input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects, false, "Use String type instead of an exception in case of ambiguous paths in JSON objects during named tuples inference", 0) \
M(Bool, input_format_json_infer_incomplete_types_as_strings, true, "Use type String for keys that contains only Nulls or empty objects/arrays during schema inference in JSON input formats", 0) \
M(Bool, input_format_json_named_tuples_as_objects, true, "Deserialize named tuple columns as JSON objects", 0) \
M(Bool, input_format_json_ignore_unknown_keys_in_named_tuple, true, "Ignore unknown keys in json object for named tuples", 0) \
@ -1025,7 +1026,7 @@ class IColumn;
M(Bool, input_format_try_infer_integers, true, "Try to infer integers instead of floats while schema inference in text formats", 0) \
M(Bool, input_format_try_infer_dates, true, "Try to infer dates from string fields while schema inference in text formats", 0) \
M(Bool, input_format_try_infer_datetimes, true, "Try to infer datetimes from string fields while schema inference in text formats", 0) \
M(Bool, input_format_try_infer_exponent_floats, false, "Try to infer floats in exponential notation while schema inference in text formats", 0) \
M(Bool, input_format_try_infer_exponent_floats, false, "Try to infer floats in exponential notation while schema inference in text formats (except JSON, where exponent numbers are always inferred)", 0) \
M(Bool, output_format_markdown_escape_special_characters, false, "Escape special characters in Markdown", 0) \
M(Bool, input_format_protobuf_flatten_google_wrappers, false, "Enable Google wrappers for regular non-nested columns, e.g. google.protobuf.StringValue 'str' for String column 'str'. For Nullable columns empty wrappers are recognized as defaults, and missing as nulls", 0) \
M(Bool, output_format_protobuf_nullables_with_google_wrappers, false, "When serializing Nullable columns with Google wrappers, serialize default values as empty wrappers. If turned off, default and null values are not serialized", 0) \

View File

@ -85,6 +85,9 @@ namespace SettingsChangesHistory
/// It's used to implement `compatibility` setting (see https://github.com/ClickHouse/ClickHouse/issues/35972)
static std::map<ClickHouseVersion, SettingsChangesHistory::SettingsChanges> settings_changes_history =
{
{"24.3", {
{"input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects", false, false, "Allow to use String type for ambiguous paths during named tuple inference from JSON objects"},
}},
{"24.2", {
{"allow_suspicious_variant_types", true, false, "Don't allow creating Variant type with suspicious variants by default"},
{"validate_experimental_and_suspicious_types_inside_nested_types", false, true, "Validate usage of experimental and suspicious types inside nested types"},

View File

@ -450,8 +450,10 @@ String getAdditionalFormatInfoByEscapingRule(const FormatSettings & settings, Fo
break;
case FormatSettings::EscapingRule::JSON:
result += fmt::format(
", try_infer_numbers_from_strings={}, read_bools_as_numbers={}, read_bools_as_strings={}, read_objects_as_strings={}, read_numbers_as_strings={}, "
"read_arrays_as_strings={}, try_infer_objects_as_tuples={}, infer_incomplete_types_as_strings={}, try_infer_objects={}",
", try_infer_numbers_from_strings={}, read_bools_as_numbers={}, read_bools_as_strings={}, read_objects_as_strings={}, "
"read_numbers_as_strings={}, "
"read_arrays_as_strings={}, try_infer_objects_as_tuples={}, infer_incomplete_types_as_strings={}, try_infer_objects={}, "
"use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects={}",
settings.json.try_infer_numbers_from_strings,
settings.json.read_bools_as_numbers,
settings.json.read_bools_as_strings,
@ -460,7 +462,8 @@ String getAdditionalFormatInfoByEscapingRule(const FormatSettings & settings, Fo
settings.json.read_arrays_as_strings,
settings.json.try_infer_objects_as_tuples,
settings.json.infer_incomplete_types_as_strings,
settings.json.allow_object_type);
settings.json.allow_object_type,
settings.json.use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects);
break;
default:
break;

View File

@ -105,6 +105,7 @@ FormatSettings getFormatSettings(const ContextPtr & context, const Settings & se
format_settings.json.write_named_tuples_as_objects = settings.output_format_json_named_tuples_as_objects;
format_settings.json.skip_null_value_in_named_tuples = settings.output_format_json_skip_null_value_in_named_tuples;
format_settings.json.read_named_tuples_as_objects = settings.input_format_json_named_tuples_as_objects;
format_settings.json.use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects = settings.input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects;
format_settings.json.defaults_for_missing_elements_in_named_tuple = settings.input_format_json_defaults_for_missing_elements_in_named_tuple;
format_settings.json.ignore_unknown_keys_in_named_tuple = settings.input_format_json_ignore_unknown_keys_in_named_tuple;
format_settings.json.quote_64bit_integers = settings.output_format_json_quote_64bit_integers;

View File

@ -202,6 +202,7 @@ struct FormatSettings
bool quote_decimals = false;
bool escape_forward_slashes = true;
bool read_named_tuples_as_objects = false;
bool use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects = false;
bool write_named_tuples_as_objects = false;
bool skip_null_value_in_named_tuples = false;
bool defaults_for_missing_elements_in_named_tuple = false;

View File

@ -136,7 +136,7 @@ namespace
bool empty() const { return paths.empty(); }
DataTypePtr finalize() const
DataTypePtr finalize(bool use_string_type_for_ambiguous_paths = false) const
{
if (paths.empty())
throw Exception(ErrorCodes::ONLY_NULLS_WHILE_READING_SCHEMA, "Cannot infer named Tuple from JSON object because object is empty");
@ -167,7 +167,7 @@ namespace
current_node->leaf_type = type;
}
return root_node.getType();
return root_node.getType(use_string_type_for_ambiguous_paths);
}
private:
@ -180,7 +180,7 @@ namespace
/// Store path to this node for better exception message in case of ambiguous paths.
String path;
DataTypePtr getType() const
DataTypePtr getType(bool use_string_type_for_ambiguous_paths) const
{
/// Check if we have ambiguous paths.
/// For example:
@ -191,7 +191,16 @@ namespace
/// And after merge we will have ambiguous paths 'a.b.c' : Int32 and 'a.b' : Nullable(Nothing),
/// but it's a valid case and we should ignore path 'a.b'.
if (leaf_type && !isNothing(removeNullable(leaf_type)) && !nodes.empty())
throw Exception(ErrorCodes::INCORRECT_DATA, "JSON objects have ambiguous paths: '{}' with type {} and '{}'", path, leaf_type->getName(), nodes.begin()->second.path);
{
if (use_string_type_for_ambiguous_paths)
return std::make_shared<DataTypeString>();
throw Exception(
ErrorCodes::INCORRECT_DATA,
"JSON objects have ambiguous paths: '{}' (with type {}) and '{}'. You can enable setting "
"input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects to use String type "
"for path '{}'",
path, leaf_type->getName(), nodes.begin()->second.path, path);
}
if (nodes.empty())
return leaf_type;
@ -203,7 +212,7 @@ namespace
for (const auto & [name, node] : nodes)
{
node_names.push_back(name);
node_types.push_back(node.getType());
node_types.push_back(node.getType(use_string_type_for_ambiguous_paths));
}
return std::make_shared<DataTypeTuple>(std::move(node_types), std::move(node_names));
@ -866,13 +875,15 @@ namespace
return std::make_shared<DataTypeTuple>(nested_types);
}
template <bool is_json>
bool tryReadFloat(Float64 & value, ReadBuffer & buf, const FormatSettings & settings)
{
if (settings.try_infer_exponent_floats)
if (is_json || settings.try_infer_exponent_floats)
return tryReadFloatText(value, buf);
return tryReadFloatTextNoExponent(value, buf);
}
template <bool is_json>
DataTypePtr tryInferNumber(ReadBuffer & buf, const FormatSettings & settings)
{
if (buf.eof())
@ -911,7 +922,7 @@ namespace
buf.position() = number_start;
}
if (tryReadFloat(tmp_float, buf, settings))
if (tryReadFloat<is_json>(tmp_float, buf, settings))
{
if (read_int && buf.position() == int_end)
return std::make_shared<DataTypeInt64>();
@ -945,7 +956,7 @@ namespace
peekable_buf.rollbackToCheckpoint(true);
}
if (tryReadFloat(tmp_float, peekable_buf, settings))
if (tryReadFloat<is_json>(tmp_float, peekable_buf, settings))
{
/// Float parsing reads no fewer bytes than integer parsing,
/// so position of the buffer is either the same, or further.
@ -957,7 +968,7 @@ namespace
return std::make_shared<DataTypeFloat64>();
}
}
else if (tryReadFloat(tmp_float, buf, settings))
else if (tryReadFloat<is_json>(tmp_float, buf, settings))
{
return std::make_shared<DataTypeFloat64>();
}
@ -966,6 +977,36 @@ namespace
return nullptr;
}
template <bool is_json>
DataTypePtr tryInferNumberFromStringImpl(std::string_view field, const FormatSettings & settings)
{
ReadBufferFromString buf(field);
if (settings.try_infer_integers)
{
Int64 tmp_int;
if (tryReadIntText(tmp_int, buf) && buf.eof())
return std::make_shared<DataTypeInt64>();
/// We can safely get back to the start of buffer, because we read from a string and we didn't reach eof.
buf.position() = buf.buffer().begin();
/// In case of Int64 overflow, try to infer UInt64
UInt64 tmp_uint;
if (tryReadIntText(tmp_uint, buf) && buf.eof())
return std::make_shared<DataTypeUInt64>();
}
/// We can safely get back to the start of buffer, because we read from a string and we didn't reach eof.
buf.position() = buf.buffer().begin();
Float64 tmp;
if (tryReadFloat<is_json>(tmp, buf, settings) && buf.eof())
return std::make_shared<DataTypeFloat64>();
return nullptr;
}
template <bool is_json>
DataTypePtr tryInferString(ReadBuffer & buf, const FormatSettings & settings, JSONInferenceInfo * json_info)
{
@ -995,7 +1036,7 @@ namespace
{
if (settings.json.try_infer_numbers_from_strings)
{
if (auto number_type = tryInferNumberFromString(field, settings))
if (auto number_type = tryInferNumberFromStringImpl<true>(field, settings))
{
json_info->numbers_parsed_from_json_strings.insert(number_type.get());
return number_type;
@ -1238,7 +1279,7 @@ namespace
}
/// Number
return tryInferNumber(buf, settings);
return tryInferNumber<is_json>(buf, settings);
}
}
@ -1294,7 +1335,7 @@ void transformFinalInferredJSONTypeIfNeededImpl(DataTypePtr & data_type, const F
return;
}
data_type = json_paths->finalize();
data_type = json_paths->finalize(settings.json.use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects);
transformFinalInferredJSONTypeIfNeededImpl(data_type, settings, json_info, remain_nothing_types);
return;
}
@ -1377,31 +1418,7 @@ void transformFinalInferredJSONTypeIfNeeded(DataTypePtr & data_type, const Forma
DataTypePtr tryInferNumberFromString(std::string_view field, const FormatSettings & settings)
{
ReadBufferFromString buf(field);
if (settings.try_infer_integers)
{
Int64 tmp_int;
if (tryReadIntText(tmp_int, buf) && buf.eof())
return std::make_shared<DataTypeInt64>();
/// We can safely get back to the start of buffer, because we read from a string and we didn't reach eof.
buf.position() = buf.buffer().begin();
/// In case of Int64 overflow, try to infer UInt64
UInt64 tmp_uint;
if (tryReadIntText(tmp_uint, buf) && buf.eof())
return std::make_shared<DataTypeUInt64>();
}
/// We can safely get back to the start of buffer, because we read from a string and we didn't reach eof.
buf.position() = buf.buffer().begin();
Float64 tmp;
if (tryReadFloat(tmp, buf, settings) && buf.eof())
return std::make_shared<DataTypeFloat64>();
return nullptr;
return tryInferNumberFromStringImpl<false>(field, settings);
}
DataTypePtr tryInferDateOrDateTimeFromString(std::string_view field, const FormatSettings & settings)

View File

@ -1,2 +1,3 @@
c1 Nullable(String)
c1 Nullable(Float64)
x Nullable(Float64)

View File

@ -1,2 +1,5 @@
DESC format(CSV, '1E20\n1.1E20') settings input_format_try_infer_exponent_floats = 0;
DESC format(CSV, '1E20\n1.1E20') settings input_format_try_infer_exponent_floats = 1;
-- This setting should not take affect on JSON formats
DESC format(JSONEachRow, '{"x" : 1.1e20}') settings input_format_try_infer_exponent_floats = 0;

View File

@ -0,0 +1,3 @@
obj Tuple(\n a Nullable(String))
('42')
('{"b" : 42}')

View File

@ -0,0 +1,4 @@
set input_format_json_use_string_type_for_ambiguous_paths_in_named_tuples_inference_from_objects=1;
desc format(JSONEachRow, '{"obj" : {"a" : 42}}, {"obj" : {"a" : {"b" : 42}}}');
select * from format(JSONEachRow, '{"obj" : {"a" : 42}}, {"obj" : {"a" : {"b" : 42}}}');