ClickHouse/programs/library-bridge/CatBoostLibraryHandler.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

392 lines
15 KiB
C++
Raw Permalink Normal View History

feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
#include "CatBoostLibraryHandler.h"
#include <Columns/ColumnTuple.h>
#include <Common/FieldVisitorConvertToNumber.h>
namespace DB
{
namespace ErrorCodes
{
extern const int BAD_ARGUMENTS;
extern const int CANNOT_APPLY_CATBOOST_MODEL;
extern const int CANNOT_LOAD_CATBOOST_MODEL;
extern const int LOGICAL_ERROR;
}
CatBoostLibraryHandler::APIHolder::APIHolder(SharedLibrary & lib)
{
ModelCalcerCreate = lib.get<CatBoostLibraryAPI::ModelCalcerCreateFunc>(CatBoostLibraryAPI::ModelCalcerCreateName);
ModelCalcerDelete = lib.get<CatBoostLibraryAPI::ModelCalcerDeleteFunc>(CatBoostLibraryAPI::ModelCalcerDeleteName);
GetErrorString = lib.get<CatBoostLibraryAPI::GetErrorStringFunc>(CatBoostLibraryAPI::GetErrorStringName);
LoadFullModelFromFile = lib.get<CatBoostLibraryAPI::LoadFullModelFromFileFunc>(CatBoostLibraryAPI::LoadFullModelFromFileName);
CalcModelPredictionFlat = lib.get<CatBoostLibraryAPI::CalcModelPredictionFlatFunc>(CatBoostLibraryAPI::CalcModelPredictionFlatName);
CalcModelPrediction = lib.get<CatBoostLibraryAPI::CalcModelPredictionFunc>(CatBoostLibraryAPI::CalcModelPredictionName);
CalcModelPredictionWithHashedCatFeatures = lib.get<CatBoostLibraryAPI::CalcModelPredictionWithHashedCatFeaturesFunc>(CatBoostLibraryAPI::CalcModelPredictionWithHashedCatFeaturesName);
GetStringCatFeatureHash = lib.get<CatBoostLibraryAPI::GetStringCatFeatureHashFunc>(CatBoostLibraryAPI::GetStringCatFeatureHashName);
GetIntegerCatFeatureHash = lib.get<CatBoostLibraryAPI::GetIntegerCatFeatureHashFunc>(CatBoostLibraryAPI::GetIntegerCatFeatureHashName);
GetFloatFeaturesCount = lib.get<CatBoostLibraryAPI::GetFloatFeaturesCountFunc>(CatBoostLibraryAPI::GetFloatFeaturesCountName);
GetCatFeaturesCount = lib.get<CatBoostLibraryAPI::GetCatFeaturesCountFunc>(CatBoostLibraryAPI::GetCatFeaturesCountName);
GetTreeCount = lib.tryGet<CatBoostLibraryAPI::GetTreeCountFunc>(CatBoostLibraryAPI::GetTreeCountName);
GetDimensionsCount = lib.tryGet<CatBoostLibraryAPI::GetDimensionsCountFunc>(CatBoostLibraryAPI::GetDimensionsCountName);
}
CatBoostLibraryHandler::CatBoostLibraryHandler(
const std::string & library_path,
const std::string & model_path)
: loading_start_time(std::chrono::system_clock::now())
, library(std::make_shared<SharedLibrary>(library_path))
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
, api(*library)
{
model_calcer_handle = api.ModelCalcerCreate();
if (!api.LoadFullModelFromFile(model_calcer_handle, model_path.c_str()))
{
throw Exception(ErrorCodes::CANNOT_LOAD_CATBOOST_MODEL,
"Cannot load CatBoost model: {}", api.GetErrorString());
}
float_features_count = api.GetFloatFeaturesCount(model_calcer_handle);
cat_features_count = api.GetCatFeaturesCount(model_calcer_handle);
tree_count = 1;
if (api.GetDimensionsCount)
tree_count = api.GetDimensionsCount(model_calcer_handle);
loading_duration = std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::system_clock::now() - loading_start_time);
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
}
CatBoostLibraryHandler::~CatBoostLibraryHandler()
{
api.ModelCalcerDelete(model_calcer_handle);
}
std::chrono::system_clock::time_point CatBoostLibraryHandler::getLoadingStartTime() const
{
return loading_start_time;
}
std::chrono::milliseconds CatBoostLibraryHandler::getLoadingDuration() const
{
return loading_duration;
}
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
namespace
{
/// Buffer should be allocated with features_count * column->size() elements.
/// Place column elements in positions buffer[0], buffer[features_count], ... , buffer[size * features_count]
template <typename T>
void placeColumnAsNumber(const IColumn * column, T * buffer, size_t features_count)
{
size_t size = column->size();
FieldVisitorConvertToNumber<T> visitor;
for (size_t i = 0; i < size; ++i)
{
/// TODO: Replace with column visitor.
Field field;
column->get(i, field);
*buffer = applyVisitor(visitor, field);
buffer += features_count;
}
}
/// Buffer should be allocated with features_count * column->size() elements.
/// Place string pointers in positions buffer[0], buffer[features_count], ... , buffer[size * features_count]
void placeStringColumn(const ColumnString & column, const char ** buffer, size_t features_count)
{
size_t size = column.size();
for (size_t i = 0; i < size; ++i)
{
2022-09-17 03:34:18 +00:00
*buffer = const_cast<char *>(column.getDataAt(i).data);
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
buffer += features_count;
}
}
/// Buffer should be allocated with features_count * column->size() elements.
/// Place string pointers in positions buffer[0], buffer[features_count], ... , buffer[size * features_count]
/// Returns PODArray which holds data (because ColumnFixedString doesn't store terminating zero).
PODArray<char> placeFixedStringColumn(const ColumnFixedString & column, const char ** buffer, size_t features_count)
{
size_t size = column.size();
size_t str_size = column.getN();
PODArray<char> data(size * (str_size + 1));
char * data_ptr = data.data();
for (size_t i = 0; i < size; ++i)
{
auto ref = column.getDataAt(i);
memcpy(data_ptr, ref.data, ref.size);
data_ptr[ref.size] = 0;
*buffer = data_ptr;
data_ptr += ref.size + 1;
buffer += features_count;
}
return data;
}
/// Place columns into buffer, returns column which holds placed data. Buffer should contains column->size() values.
template <typename T>
ColumnPtr placeNumericColumns(const ColumnRawPtrs & columns, size_t offset, size_t size, const T** buffer)
{
if (size == 0)
return nullptr;
size_t column_size = columns[offset]->size();
auto data_column = ColumnVector<T>::create(size * column_size);
T * data = data_column->getData().data();
for (size_t i = 0; i < size; ++i)
{
const auto * column = columns[offset + i];
if (column->isNumeric())
placeColumnAsNumber(column, data + i, size);
}
for (size_t i = 0; i < column_size; ++i)
{
*buffer = data;
++buffer;
data += size;
}
return data_column;
}
/// Place columns into buffer, returns data which was used for fixed string columns.
/// Buffer should contains column->size() values, each value contains size strings.
std::vector<PODArray<char>> placeStringColumns(const ColumnRawPtrs & columns, size_t offset, size_t size, const char ** buffer)
{
if (size == 0)
return {};
std::vector<PODArray<char>> data;
for (size_t i = 0; i < size; ++i)
{
const auto * column = columns[offset + i];
if (const auto * column_string = typeid_cast<const ColumnString *>(column))
placeStringColumn(*column_string, buffer + i, size);
else if (const auto * column_fixed_string = typeid_cast<const ColumnFixedString *>(column))
data.push_back(placeFixedStringColumn(*column_fixed_string, buffer + i, size));
else
throw Exception(ErrorCodes::LOGICAL_ERROR, "Cannot place string column.");
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
}
return data;
}
/// buffer[column_size * cat_features_count] -> char * => cat_features[column_size][cat_features_count] -> char *
void fillCatFeaturesBuffer(
const char *** cat_features, const char ** buffer,
size_t column_size, size_t cat_features_count)
{
for (size_t i = 0; i < column_size; ++i)
{
*cat_features = buffer;
++cat_features;
buffer += cat_features_count;
}
}
/// Calc hash for string cat feature at ps positions.
template <typename Column>
void calcStringHashes(const Column * column, size_t ps, const int ** buffer, const CatBoostLibraryHandler::APIHolder & api)
{
size_t column_size = column->size();
for (size_t j = 0; j < column_size; ++j)
{
auto ref = column->getDataAt(j);
const_cast<int *>(*buffer)[ps] = api.GetStringCatFeatureHash(ref.data, ref.size);
++buffer;
}
}
/// Calc hash for int cat feature at ps position. Buffer at positions ps should contains unhashed values.
void calcIntHashes(size_t column_size, size_t ps, const int ** buffer, const CatBoostLibraryHandler::APIHolder & api)
{
for (size_t j = 0; j < column_size; ++j)
{
const_cast<int *>(*buffer)[ps] = api.GetIntegerCatFeatureHash((*buffer)[ps]);
++buffer;
}
}
/// buffer contains column->size() rows and size columns.
/// For int cat features calc hash inplace.
/// For string cat features calc hash from column rows.
void calcHashes(const ColumnRawPtrs & columns, size_t offset, size_t size, const int ** buffer, const CatBoostLibraryHandler::APIHolder & api)
{
if (size == 0)
return;
size_t column_size = columns[offset]->size();
std::vector<PODArray<char>> data;
for (size_t i = 0; i < size; ++i)
{
const auto * column = columns[offset + i];
if (const auto * column_string = typeid_cast<const ColumnString *>(column))
calcStringHashes(column_string, i, buffer, api);
else if (const auto * column_fixed_string = typeid_cast<const ColumnFixedString *>(column))
calcStringHashes(column_fixed_string, i, buffer, api);
else
calcIntHashes(column_size, i, buffer, api);
}
}
}
/// Convert values to row-oriented format and call evaluation function from CatBoost wrapper api.
/// * CalcModelPredictionFlat if no cat features
/// * CalcModelPrediction if all cat features are strings
/// * CalcModelPredictionWithHashedCatFeatures if has int cat features.
ColumnFloat64::MutablePtr CatBoostLibraryHandler::evalImpl(
const ColumnRawPtrs & columns,
bool cat_features_are_strings) const
{
size_t column_size = columns.front()->size();
auto result = ColumnFloat64::create(column_size * tree_count);
auto * result_buf = result->getData().data();
if (!column_size)
return result;
/// Prepare float features.
PODArray<const float *> float_features(column_size);
auto * float_features_buf = float_features.data();
/// Store all float data into single column. float_features is a list of pointers to it.
auto float_features_col = placeNumericColumns<float>(columns, 0, float_features_count, float_features_buf);
if (cat_features_count == 0)
{
if (!api.CalcModelPredictionFlat(model_calcer_handle, column_size,
float_features_buf, float_features_count,
result_buf, column_size * tree_count))
{
2023-01-24 14:29:19 +00:00
throw Exception(ErrorCodes::CANNOT_APPLY_CATBOOST_MODEL,
"Error occurred while applying CatBoost model: {}", api.GetErrorString());
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
}
return result;
}
/// Prepare cat features.
if (cat_features_are_strings)
{
/// cat_features_holder stores pointers to ColumnString data or fixed_strings_data.
PODArray<const char *> cat_features_holder(cat_features_count * column_size);
PODArray<const char **> cat_features(column_size);
auto * cat_features_buf = cat_features.data();
fillCatFeaturesBuffer(cat_features_buf, cat_features_holder.data(), column_size, cat_features_count);
/// Fixed strings are stored without termination zero, so have to copy data into fixed_strings_data.
auto fixed_strings_data = placeStringColumns(columns, float_features_count,
cat_features_count, cat_features_holder.data());
if (!api.CalcModelPrediction(model_calcer_handle, column_size,
float_features_buf, float_features_count,
cat_features_buf, cat_features_count,
result_buf, column_size * tree_count))
{
2023-01-24 14:29:19 +00:00
throw Exception(ErrorCodes::CANNOT_APPLY_CATBOOST_MODEL,
"Error occurred while applying CatBoost model: {}", api.GetErrorString());
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
}
}
else
{
PODArray<const int *> cat_features(column_size);
auto * cat_features_buf = cat_features.data();
auto cat_features_col = placeNumericColumns<int>(columns, float_features_count,
cat_features_count, cat_features_buf);
calcHashes(columns, float_features_count, cat_features_count, cat_features_buf, api);
if (!api.CalcModelPredictionWithHashedCatFeatures(
model_calcer_handle, column_size,
float_features_buf, float_features_count,
cat_features_buf, cat_features_count,
result_buf, column_size * tree_count))
{
2023-01-24 14:29:19 +00:00
throw Exception(ErrorCodes::CANNOT_APPLY_CATBOOST_MODEL,
"Error occurred while applying CatBoost model: {}", api.GetErrorString());
feat: implement catboost in library-bridge This commit moves the catboost model evaluation out of the server process into the library-bridge binary. This serves two goals: On the one hand, crashes / memory corruptions of the catboost library no longer affect the server. On the other hand, we can forbid loading dynamic libraries in the server (catboost was the last consumer of this functionality), thus improving security. SQL syntax: SELECT catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction, ACTION AS target FROM amazon_train LIMIT 10 Required configuration: <catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path> *** Implementation Details *** The internal protocol between the server and the library-bridge is simple: - HTTP GET on path "/extdict_ping": A ping, used during the handshake to check if the library-bridge runs. - HTTP POST on path "extdict_request" (1) Send a "catboost_GetTreeCount" request from the server to the bridge, containing a library path (e.g /home/user/libcatboost.so) and a model path (e.g. /home/user/model.bin). Rirst, this unloads the catboost library handler associated to the model path (if it was loaded), then loads the catboost library handler associated to the model path, then executes GetTreeCount() on the library handler and finally sends the result back to the server. Step (1) is called once by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The library path handler is unloaded in the beginning because it contains state which may no longer be valid if the user runs catboost("/path/to/model.bin", ...) more than once and if "model.bin" was updated in between. (2) Send "catboost_Evaluate" from the server to the bridge, containing the model path and the features to run the interference on. Step (2) is called multiple times (once per chunk) by the server from function FunctionCatBoostEvaluate::executeImpl(). The library handler for the given model path is expected to be already loaded by Step (1). Fixes #27870
2022-08-05 07:53:06 +00:00
}
}
return result;
}
size_t CatBoostLibraryHandler::getTreeCount() const
{
std::lock_guard lock(mutex);
return tree_count;
}
ColumnPtr CatBoostLibraryHandler::evaluate(const ColumnRawPtrs & columns) const
{
std::lock_guard lock(mutex);
if (columns.empty())
throw Exception(ErrorCodes::BAD_ARGUMENTS, "Got empty columns list for CatBoost model.");
if (columns.size() != float_features_count + cat_features_count)
throw Exception(ErrorCodes::BAD_ARGUMENTS,
"Number of columns is different with number of features: columns size {} float features size {} + cat features size {}",
columns.size(),
float_features_count,
cat_features_count);
for (size_t i = 0; i < float_features_count; ++i)
{
if (!columns[i]->isNumeric())
{
throw Exception(ErrorCodes::BAD_ARGUMENTS, "Column {} should be numeric to make float feature.", i);
}
}
bool cat_features_are_strings = true;
for (size_t i = float_features_count; i < float_features_count + cat_features_count; ++i)
{
const auto * column = columns[i];
if (column->isNumeric())
{
cat_features_are_strings = false;
}
else if (!(typeid_cast<const ColumnString *>(column)
|| typeid_cast<const ColumnFixedString *>(column)))
{
throw Exception(ErrorCodes::BAD_ARGUMENTS, "Column {} should be numeric or string.", i);
}
}
auto result = evalImpl(columns, cat_features_are_strings);
if (tree_count == 1)
return result;
size_t column_size = columns.front()->size();
auto * result_buf = result->getData().data();
/// Multiple trees case. Copy data to several columns.
MutableColumns mutable_columns(tree_count);
std::vector<Float64 *> column_ptrs(tree_count);
for (size_t i = 0; i < tree_count; ++i)
{
auto col = ColumnFloat64::create(column_size);
column_ptrs[i] = col->getData().data();
mutable_columns[i] = std::move(col);
}
Float64 * data = result_buf;
for (size_t row = 0; row < column_size; ++row)
{
for (size_t i = 0; i < tree_count; ++i)
{
*column_ptrs[i] = *data;
++column_ptrs[i];
++data;
}
}
return ColumnTuple::create(std::move(mutable_columns));
}
}