Native DB4AI Engine

The current version of openGauss supports the native DB4AI capability. By introducing native AI operators, openGauss simplifies the operation process and fully utilizes the optimization and execution capabilities of the database optimizer and executor to obtain the high-performance model training capability in the database. With a simpler model training and prediction process and higher performance, developers can focus on model tuning and data analysis in a shorter period of time, avoiding fragmented technology stacks and redundant code implementation.

Keyword Parsing

Table 1 DB4AI syntax and keywords

Name		Description
Syntax	CREATE MODEL	Creates a model, trains it, and saves the model.
	PREDICT BY	Uses an existing model for prediction.
	DROP MODEL	Deletes a model.
Keyword	TARGET	Target column name of a training or prediction task.
	FEATURES	Data feature column name of a training or prediction task.
	MODEL	Model name of a training task.

Usage Guide

Introduce the algorithms supported in this version.

The DB4AI of the current version supports the following new algorithms:

Table 2 Supported algorithms

Optimization Algorithm	Algorithm
GD	logistic_regression
	linear_regression
	svm_classification
	PCA
	multiclass
Kmeans	kmeans
xgboost	xgboost_regression_logistic
	xgboost_binary_logistic
	xgboost_regression_squarederror
	xgboost_regression_gamma

Learn about the model training syntax.

CREATE MODEL

You can run the CREATE MODEL statement to create and train a model. This SQL statement uses the public Iris dataset for model training.

The following uses multiclass as an example to describe how to train a model. Specify sepal_length, sepal_width, petal_length, and petal_width as feature columns in the tb_iris training set, and use the multiclass algorithm to create and save the iris_classification_model model.

MogDB=# CREATE MODEL iris_classification_model USING xgboost_regression_logistic FEATURES sepal_length, sepal_width,petal_length,petal_width TARGET target_type < 2 FROM tb_iris_1 WITH nthread=4, max_depth=8;
MODEL CREATED. PROCESSED 1

In the preceding command:

The CREATE MODEL statement is used to train and save a model.
USING specifies the algorithm name.
FEATURES specifies the features of the training model and needs to be added based on the column name of the training data table.
TARGET specifies the training target of the model. It can be the column name of the data table required for training or an expression, for example, price > 10000.
WITH specifies the hyperparameters used for model training. When the hyperparameter is not set by the user, the framework uses the default value.

The framework supports various hyperparameter combinations for different operators.

Table 3 Hyperparameters supported by operators

Operator	Hyperparameter
GD (logistic_regression、linear_regression、svm_classification)	optimizer(char), verbose(bool), max_iterations(int), max_seconds(double), batch_size(int), learning_rate(double), decay(double), and tolerance(double) SVM limits the hyperparameter lambda(double).
K-Means	max_iterations(int), num_centroids(int), tolerance(double), batch_size(int), num_features(int), distance_function(char), seeding_function(char), verbose(int), and seed(int)
GD(pca)	batch_size(int);max_iterations(int);max_seconds(int);tolerance(float8);verbose(bool);number_components(int);seed(int)
GD(multiclass)	classifier(char) Note: Other hyperparameter types of multiclass depend on the categories in the selected classifier.
xgboost_regression_logistic、xgboost_binary_logistic、xgboost_regression_squarederror、xgboost_regression_gamma	batch_size(int);booster(char);tree_method(char);eval_metric(char);seed(int);nthread(int);max_depth(int);gamma(float8);eta(float8);min_child_weight(int);verbosity(int)

The default value and value range of each hyperparameter are as follows:

Table 4 Default values and value ranges of hyperparameters

Operator	Default Hyperparameter Value	Value Range	Hyperparameter Description
GD:logistic_regression、linear_regression、svm_classification、pca	optimizer = gd (gradient descent)	gd/ngd (natural gradient descent)	Optimizer
	verbose = false	T/F	Log display
	max_iterations = 100	(0, 10000]	Maximum iterations
	max_seconds = 0 (The running duration is not limited.)	[0,INT_MAX_VALUE]	Running duration
	batch_size = 1000	(0, 1048575]	Number of data records selected per training
	learning_rate = 0.8	(0, DOUBLE_MAX_VALUE]	Learning rate
	decay = 0.95	(0, DOUBLE_MAX_VALUE]	Weight decay rate
	tolerance = 0.0005	(0, DOUBLE_MAX_VALUE]	Tolerance
	seed = 0 (random value of seed)	[0, INT_MAX_VALUE]	Seed
	just for linear、SVM：kernel = “linear”	linear/gaussian/polynomial	Kernel function
	just for linear、SVM：components = MAX(2*features, 128)	[0, INT_MAX_VALUE]	Number of high-dimension space dimensions
	just for linear、SVM：gamma = 0.5	(0, DOUBLE_MAX_VALUE]	Gaussian kernel function parameter
	just for linear、SVM：degree = 2	[2, 9]	Polynomial kernel function parameter
	just for linear、SVM：coef0 = 1.0	[0, DOUBLE_MAX_VALUE]	Polynomial kernel function parameter
	just for SVM：lambda = 0.01	(0, DOUBLE_MAX_VALUE)	Regularization parameter
	just for pca： number_components	（0，INT_MAX_VALUE]	Target dimension after dimension reduction
GD:multiclass	classifier=“svm_classification”	svm_classification\logistic_regression	Classifier for multiclass tasks
Kmeans	max_iterations = 10	[1, 10000]	Maximum iterations
	num_centroids = 10	[1, 1000000]	Number of clusters
	tolerance = 0.00001	(0,1]	Central point error
	batch_size = 10	[1,1048575]	Number of data records selected per training
	num_features = 2	[1, INT_MAX_VALUE]	Number of sample features
	distance_function = “L2_Squared”	L1\L2\L2_Squared\Linf	Regularization method
	seeding_function = “Random++”	“Random++”\“KMeans	Method for initializing seed points
	verbose = 0U	{ 0, 1, 2 }	Verbose mode
	seed = 0U	[0, INT_MAX_VALUE]	Seed
xgboost: xgboost_regression_logistic、xgboost_binary_logistic、xgboost_regression_gamma、xgboost_regression_squarederror	n_iter=10	(0, 10000]	Iteration times
	batch_size=10000	(0, 1048575]	Number of data records selected per training
	booster=“gbtree”	gbtree\gblinear\dart	Booster type
	tree_method=“auto”	auto\exact\approx\hist\gpu_hist Note: To use the gpu_hist parameter, you must configure a GPU library. Otherwise, the DB4AI platform does not support this value.	Tree construction algorithm
	eval_metric=“rmse”	rmse\rmsle\map\mae\auc\aucpr	Data verification metric
	seed=0	[0, 100]	Seed
	nthread=1	(0, MAX_MEMORY_LIMIT]	Concurrency
	max_depth=5	(0, MAX_MEMORY_LIMIT]	Maximum depth of the tree. This parameter is valid only for the tree booster.
	gamma=0.0	[0, 1]	Minimum loss required for further partitioning on leaf nodes
	eta=0.3	[0, 1]	Step used in the update to prevent overfitting
	min_child_weight=1	[0, INT_MAX_VALUE]	Minimum sum of instance weights required by child nodes
	verbosity=1	0 (silent)\1 (warning)\2 (info)\3 (debug)	Printing level
MAX_MEMORY_LIMIT = Maximum number of tuples loaded in memory
GS_MAX_COLS = Maximum number of attributes in a database table

If the model is saved successfully, the following information is returned:
```
MODEL CREATED. PROCESSED x
```

View the model information.

After the training is complete, the model is stored in the gs_model_warehouse system catalog. You can view information about the model and training process in the gs_model_warehouse system catalog.

The model details are stored in the system catalog in binary mode. You can use the gs_explain_model function to view the model details. The statement is as follows:

MogDB=# select * from gs_explain_model("iris_classification_model");
 DB4AI MODEL
-------------------------------------------------------------
 Name: iris_classification_model
 Algorithm: xgboost_regression_logistic
 Query: CREATE MODEL iris_classification_model
 USING xgboost_regression_logistic
 FEATURES sepal_length, sepal_width,petal_length,petal_width
 TARGET target_type < 2
 FROM tb_iris_1
 WITH nthread=4, max_depth=8;
 Return type: Float64
 Pre-processing time: 0.000000
 Execution time: 0.001443
 Processed tuples: 78
 Discarded tuples: 0
 n_iter: 10
 batch_size: 10000
 max_depth: 8
 min_child_weight: 1
 gamma: 0.0000000000
 eta: 0.3000000000
 nthread: 4
 verbosity: 1
 seed: 0
 booster: gbtree
 tree_method: auto
 eval_metric: rmse
 rmse: 0.2648450136
 model size: 4613

Use an existing model to perform a prediction task.

Use the SELECT and PREDICT BY keywords to complete the prediction task based on the existing model.

Query syntax: SELECT… PREDICT BY… (FEATURES…)… FROM…;
```
MogDB=# SELECT id, PREDICT BY iris_classification (FEATURES sepal_length,sepal_width,petal_length,petal_width) as "PREDICT" FROM tb_iris limit 3;

id  | PREDICT
-----+---------
  84 |       2
  85 |       0
  86 |       0
(3 rows)
```
For the same prediction task, the results of the same model are stable. In addition, models trained based on the same hyperparameter and training set are stable. AI model training is random (random gradient descent of data distribution each batch). Therefore, the computing performance and results of different models can vary slightly.

View the execution plan.

You can use the EXPLAIN statement to analyze the execution plan in the model training or prediction process of CREATE MODEL and PREDICT BY. The keyword EXPLAIN can be followed by a CREATE MODEL or PREDICT BY clause or an optional parameter. The supported parameters are as follows:

Table 5 Parameters supported by EXPLAIN

Parameter	Description
ANALYZE	Boolean variable, which is used to add description information such as the running time and number of loop times
VERBOSE	Boolean variable, which determines whether to output the training running information to the client
COSTS	Boolean variable
CPU	Boolean variable
DETAIL	Boolean variable, which is unavailable
NODES	Boolean variable, which is unavailable
NUM_NODES	Boolean variable, which is unavailable
BUFFERS	Boolean variable
TIMING	Boolean variable
PLAN	Boolean variable
FORMAT	Optional format type: TEXT, XML, JSON, and YAML

Example:

MogDB=# Explain CREATE MODEL patient_logisitic_regression USING logistic_regression FEATURES second_attack, treatment TARGET trait_anxiety > 50 FROM patients WITH batch_size=10, learning_rate = 0.05;
                               QUERY PLAN
-------------------------------------------------------------------------
 Train Model - logistic_regression  (cost=0.00..0.00 rows=0 width=0)
   ->  Materialize  (cost=0.00..41.08 rows=1776 width=12)
         ->  Seq Scan on patients  (cost=0.00..32.20 rows=1776 width=12)
(3 rows)

Perform troubleshooting in case of exceptions.

Training phase

Scenario 1: When the value of the hyperparameter exceeds the value range, the model training fails and an error message is returned. For example:

MogDB=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET trait_anxiety  FROM patients WITH optimizer='aa';
ERROR:  Invalid hyperparameter value for optimizer. Valid values are: gd, ngd.

Scenario 2: If the model name already exists, the model fails to be saved, and an error message with the cause is displayed. For example:

MogDB=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET trait_anxiety  FROM patients;
ERROR:  The model name "patient_linear_regression" already exists in gs_model_warehouse.

Scenario 3: If the value in the FEATURE or TARGETS column is *****, an error message with the cause is displayed. For example:

MogDB=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES *  TARGET trait_anxiety  FROM patients;
ERROR:  FEATURES clause cannot be *
-----------------------------------------------------------------------------------------------------------------------
MogDB=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET *  FROM patients;
ERROR:  TARGET clause cannot be *

Scenario 4: If the keyword TARGET is used in the unsupervised learning method or is not applicable to the supervised learning method, an error message with the cause is displayed. For example:

MogDB=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment FROM patients;
ERROR:  Supervised ML algorithms require TARGET clause
-----------------------------------------------------------------------------------------------------------------------------
CREATE MODEL patient_linear_regression USING linear_regression TARGET trait_anxiety  FROM patients;
ERROR:  Supervised ML algorithms require FEATURES clause

Scenario 5: If there is only one category in the TARGET column, an error message with the cause is displayed. For example:

MogDB=# CREATE MODEL ecoli_svmc USING multiclass FEATURES f1, f2, f3, f4, f5, f6, f7 TARGET cat FROM (SELECT * FROM db4ai_ecoli WHERE cat='cp');
ERROR:  At least two categories are needed

Scenario 6: DB4AI filters out data that contains null values during training. When the model data involved in training is null, an error message with the cause is displayed. For example:

MogDB=# create model iris_classification_model using xgboost_regression_logistic features message_regular target error_level from error_code;
ERROR:  Training data is empty, please check the input data.

Scenario 7: The DB4AI algorithm has restrictions on the supported data types. If the data type is not in the whitelist, an error message is returned and the invalid OID is displayed. You can check the OID in pg_type to determine the invalid data type. For example:
```
MogDB=# CREATE MODEL ecoli_svmc USING multiclass FEATURES f1, f2, f3, f4, f5, f6, f7, cat TARGET cat FROM db4ai_ecoli ;
ERROR:  Oid type 1043 not yet supported
```
Scenario 8: If the GUC parameter statement_timeout is set, the statement that is executed due to training timeout will be terminated. In this case, execute the CREATE MODEL statement. Parameters such as the size of the training set, number of training rounds (iteration), early termination conditions (tolerance and max_seconds), and number of parallel threads (nthread) affect the training duration. When the duration exceeds the database limit, the statement execution is terminated and model training fails.

Model parsing
- Scenario 9: If the model name cannot be found in the system catalog, an error message with the cause is displayed. For example:
```
MogDB=# select gs_explain_model("ecoli_svmc");
ERROR:  column "ecoli_svmc" does not exist
```

Prediction phase

Scenario 10: If the model name cannot be found in the system catalog, an error message with the cause is displayed. For example:

MogDB=# select id, PREDICT BY patient_logistic_regression (FEATURES second_attack,treatment) FROM patients;
ERROR:  There is no model called "patient_logistic_regression".

Scenario 11: If the data dimension and data type of the FEATURES task are inconsistent with those of the training set, an error message with the cause is displayed. For example:

MogDB=# select id, PREDICT BY patient_linear_regression (FEATURES second_attack) FROM patients;
ERROR:  Invalid number of features for prediction, provided 1, expected 2
CONTEXT:  referenced column: patient_linear_regression_pred
-------------------------------------------------------------------------------------------------------------------------------------
MogDB=# select id, PREDICT BY patient_linear_regression (FEATURES 1,second_attack,treatment) FROM patients;
ERROR:  Invalid number of features for prediction, provided 3, expected 2
CONTEXT:  referenced column: patient_linear_regression_pre

NOTE: The DB4AI feature requires data access for computing and is not applicable to encrypted databases.

Issue