Creating a Loading Job
After a graph schema has been created, the system is ready to load data into the graph store. The GSQL language offers easy-to-understand and easy-to-use commands for data loading which perform many of the same data conversion, mapping, filtering, and merging operations that are found in enterprise ETL (Extract, Transform, and Load) systems.
The GSQL system can read structured or semistructured data from text files. The loading language syntax is geared towards tabular or JSON data, but conditional clauses and data manipulation functions allow for reading data that is structured in a more complex or irregular way. For tabular data, each line in the data file contains a series of data values, separated by commas, tabs, spaces, or any other designated ASCII characters (only single character separators are supported). A line should contain only data values and separators, without extra whitespace. From a tabular view, each line of data is a row, and each row consists of a series of column values.
Loading data is a two-step process. First, a loading job is defined. Next, the job is executed with a RUN LOADING JOB
statement. These two statements, and the components of the loading job, are detailed below.
The structure of a loading job will be presented hierarchically, top-down:
CREATE LOADING JOB, which may contain a set of DEFINE and LOAD statements
DEFINE statements
LOAD statements, which can have several clauses
All blank spaces are meaningful in string fields in CSV and JSON. Either pre-process your data files to remove extra spaces, or use GSQL's token processing functions gsql_trim
, gsql_ltrim
, and gsql_rtrim
(Built-in Loader Token Functions).
Loading job capabilities
TigerGraph's syntax for defining and running loading jobs offers several advantages:
The TigerGraph platform can handle concurrent loading jobs, which can greatly increase throughput.
The data file locations can be specified at compile time or at run time. Run-time settings override compile-time settings.
A loading job definition can include several input files. When running the job, the user can choose to run only part of the job by specifying only some of the input files.
Loading jobs can be monitored, aborted, and restarted.
Concurrent Loading
Among its several duties, the RESTPP component manages loading jobs. There can be multiple RESTPP-LOADER subcomponents, each of which can handle a loading job independently. The maximum number of concurrent loading jobs is set by the configuration parameter FileLoader.ReplicaNumber
.
Furthermore, if the TigerGraph graph is distributed (partitioned) across multiple machine nodes, each machine's RESTPP-LOADER(s) can be put into action. Each RESTPP-LOADER only reads local input data files, but the resulting graph data can be stored on any machine in the cluster.
To maximize loading performance in a cluster, use at least two loaders per machine, and assign each loader approximately the same amount of data.
A concurrent-capable loading job can logically be separated into parts according to each file variable. When a concurrent-capable loading job is compiled, a RESTPP endpoint is generated for the loading job, which you can call to load data into your graph as an alternative to RUN LOADING JOB
.
Example loading jobs and data files for the book_rating
schema defined earlier in the document are available in the $(gadmin config get System.AppRoot)/document/examples
folder in your TigerGraph platform installation.
CREATE LOADING JOB
The CREATE LOADING JOB statement is used to define a block of DEFINE, LOAD, and DELETE statements for loading data to or removing data from a particular graph. The sequence of statements is enclosed in curly braces. Each statement in the block, including the last one, should end with a semicolon.
LOAD or DELETE Statements
A loading job may contain either LOAD
or DELETE
statements but not both.
A loading job that includes both will be rejected when the CREATE
statement is executed.
Example:
DROP JOB statement
To drop (remove) a job, run DROP JOB job_name
. The job will be removed from GSQL. To drop all jobs, run either of the following commands:
The scope of ALL
depends on the user's current scope. If the user has set a working graph, then DROP ALL
removes all the jobs for that graph. If a superuser has set their scope to be global, then DROP ALL
removes all jobs across all graph spaces.
DEFINE
statements
DEFINE
statementsA DEFINE
statement is used to define a local variable or expression to be used by the subsequent LOAD
statements in the loading job.
DEFINE FILENAME
DEFINE FILENAME
The DEFINE FILENAME
statement defines a filename variable. The variable can then be used later in the JOB
block by a LOAD
statement to identify its data source. Every concurrent loading job must have at least one DEFINE FILENAME
statement.
The filevar
is optionally followed by a filepath_string
, which tells the job where to find input data. As the name suggests, filepath_string is a string value. Therefore, it should start and end with double quotes.
filepath_string
There are four options for filepath_string :
path
: either an absolute path or relative path for either a file or a folder on the machine where the job is run. If it is a folder, then the loader will attempt to load each non-hidden file in the folder.
If this path is not valid when CREATE LOADING JOB is executed, GSQL will report an error.
An absolute path may begin with the session variable $sys.data_root.
Then, when running this loading job, first set a value for the parameter, and then run the job:
As the name implies, session parameters only retain their value for the duration of the current GSQL session. If the user exits GSQL, the settings are lost.
"all:" path
: If the path is prefixed withall:
, then the loading job will attempt to run on every machine in the cluster which has a RESTPP component, and each machine will look locally for data at path. If the path is invalid on any of the machines, the job will be aborted. Also, the session parameter $sys.data_root may not be used.
"any:" path
: If the path is prefixed withany:
, then the loading job will attempt to run on every machine in the cluster which has a RESTPP component, and each machine will look locally for data at the specified path. If the path is invalid on any of the machines, those machines where the path is not valid are skipped. Also, the session parameter $sys.data_root may not be used.
A list of machine-specific paths: A machine_alias is a name such as m1, m2, etc. which is defined when the cluster configuration is set. For this option, the filepath_string may include a list of paths, separated by commas. If several machines have the same path, the paths can be grouped together by using a list of machine aliases, with the vertical bar "|" as a separator. The loading job will run on whichever machines are named; each RESTPP-LOADER will work on its local files.
DEFINE HEADER
The DEFINE HEADER statement defines a sequence of column names for an input data file. The first column name maps to the first column, the second column name maps to the second column, etc.
DEFINE INPUT_LINE_FILTER
The DEFINE INPUT_LINE_FILTER statement defines a named Boolean expression whose value depends on column attributes from a row of input data. When combined with a USING reject_line_rule clause in a LOAD statement, the filter determines whether an input line is ignored or not.
LOAD statements
A LOAD statement tells the GSQL loader how to parse a data line into column values (tokens), and then describes how the values should be used to create a new vertex or edge instance. One LOAD statement can be used to generate multiple vertices or edges, each vertex or edge having its own Destination_Clause, as shown below. Additionally, two or more LOAD statements may refer to the same input data file. In this case, the GSQL loader will merge their operations so that both of their operations are executed in a single pass through the data file.
The LOAD statement has many options. This reference guide provides examples of key features and options. The Knowledge Base and FAQs and the tutorials, such as GSQL 101, provide additional solution- and application-oriented examples.
Different LOAD statement types have different rules for the USING clause; see the USING clause section below for specifics.
LOAD statement
filevar
must have been previously defined in a DEFINE FILENAME statement.
filepath_string
must satisfy the same rules given above in the DEFINE FILENAME section.
Position-based file identifier
When a CREATE LOADING JOB
block is processed, the GSQL system will count the number of unique filepath strings (filepath_string
) and assign them position-based index numbers 0, 1, 2, etc. starting from the top. A filepath string is considered one item, even if it has multiple machine indexes and file locations. These index numbers can then be used as an alternate naming scheme for the filespath_strings.
When running a loading job, a filepath string can be referred to as __GSQL_FILENAME_n__
, where n
is replaced with the index number. For example, for the following loading job:
The filepath /home/data/person.csv
in the LOAD
statement can be referred to as __GSQL_FILENAME_0__
. While the filepath in the second LOAD
statement can be referred to as __GSQL_FILENAME_1__
.
The remainder of this section of the document will provide details on the format and use of the file_path, Destination_Clause, its subclauses. USING clause is introduced later in Section "Other Optional LOAD Clauses".
Destination Clause
A Destination Clause describes how the tokens from a data source should be used to construct one of three types of data objects: a vertex, an edge, or a row in a temporary table (TEMP_TABLE
). The destination clause formats for the three types are very similar, but we show them separately for clarity:
For the TO VERTEX
and TO EDGE
destination clauses, the vertex_type_name or edge_type_name must match the name of a vertex or edge type previously defined in a CREATE VERTEX
or CREATE UNDIRECTED|DIRECTED EDGE
statement. The values in the value list(id_expr, attr_expr1, attr_expr2,...) are assigned to the id(s) and attributes of a new vertex or edge instance, in the same order in which they are listed in the CREATE
statement. id_expr obeys the same attribute rules as attr_expr, except that only attr_expr can use the reducer function, which is introduced later.
The TO TEMP_TABLE
clause defines a new, temporary data structure. Its unique characteristics will be described in a separate subsection. For now, we focus on TO VERTEX
and TO EDGE
.
For edge clauses, the source_id_expr and target_id_expr can each optionally be followed by a source_type_expr and target_type_expr, respectively. The source_type_expr and target_type_expr must evaluate to one of the allowed endpoint vertex types for the given edge type. By specifying the vertex type, this tells the loader what id types to expect. This may be important when the edge type is defined to accept more than one type of source/target vertex.
For fast loading of edge data, referential integrity checking is disabled by default.
For an edge to be valid, it must refer to endpoint vertices that exist. To support fast, out-of-order loading, if one or both of the endpoint vertices do not yet exist, the loader will create vertices with the necessary IDs and default attribute values. Due to the loader's UPSERT semantics, if the vertex data is loaded later, it will be automatically merged with the dummy vertices. The user can disable this feature and perform regular referential integrity checking by setting the VERTEX_MUST_EXIST=true
option.
Example: Suppose we have the following vertex and edge types:
A Visit
edge can connect two Person
vertices or a Person
to a Company
. A Person
has a string ID, while a Company has an INT
ID. Then suppose the Visit
edge source data comes from a single CSV file, containing both variants of edges. Note that the 2nd column ($1) contains either Person
or Company
, and that the 3rd column ($2) contains either a string or an integer.
Using the optional target_type_expr
field, we can load both variants of the Visit
edge with a single clause.
Known issue: you must include a USING clause
when loading data into edge types with different FROM-TO
vertex pairs, even if all options are default.
Attributes and Attribute Expressions
A LOAD
statement processes each line of an input file, splitting each line (according to the SEPARATOR
character, see Section "Other Optional LOAD
Clauses" for more details) into a sequence of tokens. Each destination clause provides a token-to-attribute mapping which defines how to construct a new vertex, an edge, or a temp table row instance (e.g., one data object). The tokens can also be thought of as the column values in a table. There are two ways to refer to a column, by position or by name. Assuming a column has a name, either method may be used, and both methods may be used within one expression.
By Position: The columns (tokens) are numbered from left to right, starting with $0. The next column is $1, and so on.
By Name: Columns can be named, either through a header line in the input file, or through a DEFINE HEADER
statement. If a header line is used, then the first line of the input file should be structured like a data line, using the same separator characters, except that each column contains a column name string instead of a data value. Names are enclosed in double quotes, e.g. $"age"
.
Data file name: $sys.file_name refers to the current input data file.
In a simple case, a token value is copied directly to an attribute. For example, in the following LOAD
statement,
The
PRIMARY_ID
of a person vertex comes from column$0
of the file "xx/yy/a.csv".The next attribute of a person vertex comes from column
$1
.The next attribute of a person vertex is given the value "xx/y/a.csv" (the filename itself).
Users do not need to explicitly define a primary ID. Given the attributes, one will be selected as the primary key.
Cumulative Loading
A basic principle in the GSQL Loader is cumulative loading. Cumulative loading means that a particular data object might be written to (i.e., loaded) multiple times, and the result of the multiple loads may depend on the full sequence of writes. This usually means that If a data line provides a valid data object, and the WHERE
clause and OPTION
clause are satisfied, then the data object is loaded.
Valid input: For each input data line, each destination clause constructs one or more new data objects. To be a valid data object, it must have an ID value of the correct type, have correctly typed attribute values, and satisfy the optional
WHERE
clause. If the data object is not valid, the object is rejected (skipped) and counted as an error in the log file. The rules for invalid attributes values are summarized below:
UINT
: Any non-digit character. (Out-of-range values cause overflow instead of rejection)INT
: Any non-digit or non-sign character. (Out-of-range values cause overflow instead of rejection)FLOAT
andDOUBLE
: Any wrong formatSTRING
,STRING COMPRESS
,FIXED_BINARY
: N/ADATETIME
: Wrong format, invalid date time, or out of range.BOOL
: Any value not listed later.Complex type: Depends on the field type or element type. Any invalid field (in
UDT
), element (inLIST
orSET
), key or value (inMAP
) causes rejection.
New data objects: If a valid data object has a new ID value, then the data object is added to the graph store. Any attributes which are missing are assigned the default value for that data type or for that attribute.
Overwriting existing data objects: If a valid data object has a ID value for an existing object, then the new object overwrites the existing data object, with the following clarifications and exceptions:
The attribute values of the new object overwrite the attribute values of the existing data object.
Missing tokens: If a token is missing from the input line so that the generated attribute is missing, then that attribute retains its previous value.
A STRING token is never considered missing; if there are no characters, then the string is the empty string
Skipping an attribute: A
LOAD
statement can specify that a particular attribute should not be loaded by using the special character _ (underscore) as its attribute expression (attr_expr). For example,
means to skip the next-to-last attribute. This technique is used when it is known that the input data file does not contain data for every attribute.
If the load operation is creating a new vertex or edge, then the skipped attribute will be assigned the default value.
If the load operation is overwriting an existing vertex or edge, then the skipped attribute will retain its existing value.
More Complex Attribute Expressions
An attribute expression may use column tokens (e.g., $0
), literals (constant numeric or string values), any of the built-in loader token functions, or a user-defined token function. Attribute expressions may not contain mathematical or boolean operators (such as +
, *
, AND
). The rules for attribute expressions are the same as those for id expressions, but an attribute expression can additionally use a reducer function:
id_expr := $column_number | $"column_name" | constant | $sys.file_name | token_function_name( id_expr [, id_expr ]* )
attr_expr := id_expr | REDUCE(reducer_function_name(id _expr ))
Note that token functions can be nested, that is, a token function can be used as an input parameter for another token function. The built-in loader token/reducer functions and user-defined token functions are described in the section "Built-In Loader Token Functions".
The subsections below describe details about loading particular data types.
Loading a DOUBLE or FLOAT Attribute
A floating point value has the basic format
In the first case, the decimal point and following digits are required. In the second case, some digits are required (looking like an integer), and the following decimal point and digits are optional.
In both cases, the leading sign ( "+" or "-") is optional. The exponent, using "e" or "E", is optional. Commas and extra spaces are not allowed.
Loading a DATETIME
Attribute
DATETIME
AttributeWhen loading data into a DATETIME
attribute, the GSQL loader will automatically read a string representation of DateTime information and convert it to internal DateTime representation. The loader accepts any of the following string formats:
%Y-%m-%d %H:%M:%S
(e.g., 2011-02-03 01:02:03)%Y/%m/%d %H:%M:%S
(e.g., 2011/02/03 01:02:03)%Y-%m-%dT%H:%M:%S.000z
(e.g., 2011-02-03T01:02:03.123z, 123 will be ignored)%Y-%m-%d
(only date, no time, e.g., 2011-02-03 )%Y/%m/%d
(only date, no time, e.g., 2011/02/03)Any integer value (Unix Epoch time, where Jan 1, 1970 at 00:00:00 is integer 0)
Format notation:
%Y is a 4-digit year. A 2-digit year is not a valid value.
%m and %s are a month (1 to 12) and a day (1 to 31), respectively. Leading zeroes are optional.
%H, %M, %S are hours (0 to 23), minutes (0 to 59) and seconds (0 to 59), respectively. Leading zeroes are optional.
When loading data, the loader checks whether the values of year, month, day, hour, minute, second are out of the valid range. If any invalid value is present, e.g. '2010-13-05' or '2004-04-31 00:00:00', the attribute is invalid and the object (vertex or edge) is not created.
Loading a BOOL Attribute
When loading data from CSV files the following values are accepted for BOOL attributes :
True:
TRUE
,True
,true
,1
False:
FALSE
,False
,false
,0
When loading data from JSON documents, the valid BOOL values are true
and false
.
Loading a User-Defined Type (UDT) Attribute
To load a UDT attribute, state the name of the UDT type, followed by the list of attribute expressions for the UDT's fields, in parentheses. See the example below.
Loading a LIST or SET Attribute
There are three methods to load a LIST
or a SET
.
The first method is to load multiple rows of data that share the same ID values and append the individual attribute values to form a collection of values. The collections are formed incrementally by reading one value from each eligible data line and appending the new value into the collection. When the loading job processes a line, it checks to see whether a vertex or edge with that id value(s) already exists or not. If the id value(s) is new, then a new vertex or edge is created with a new list/set containing the single value. If the id(s) has been used before, then the value from the new line is appended to the existing list/set. Below shows an example:
The job load_set_list
will load two test_vertex
vertices because there are two unique id values in the data file. Vertex 1 has attribute values with iset = [10,20]
and ilist = [10,20,20]
. Vertex 3 has values iset = [30,40]
and ilist = [30, 30, 40]
. Note that a set doesn't contain duplicate values, while a list can contain duplicate values.
Because GSQL loading is multi-threaded, the order of values loaded into a LIST might not match the input order.
If the input file contains multiple columns which should be all added to the LIST or SET, then a second method is available. Use the LIST() or SET() function as in the example below:
The third method is to use the SPLIT()
function to read a compound token and split it into a collection of elements, to form a LIST
or SET
collection. The SPLIT()
function takes two arguments: the column index and the element separator. The element separator should be distinct from the separator throughout the whole file. Below shows an example:
The SPLIT()
function cannot be used for UDT type elements.
Loading a MAP Attribute
There are three methods to load a MAP
.
The first method is to load multiple rows of data that share the same id values. The maps are formed incrementally by reading one key-value pair from each eligible data line. When the loading job processes a line, it checks to see whether a vertex or edge with that id value(s) already exists or not. If the id value(s) is new, then a new vertex or edge is created with a new map containing the single key-value pair. If the id(s) has been used before, then the loading job checks whether the key exists in the map or not. If the key doesn't exist in the map, the new key-value pair is inserted. Otherwise, the value will be replaced by the new value.
The loading order might not be the same as the order in the raw data. If a data file contains multiple lines with the same id and same key but different values, loading them together results in a nondeterministic final value for that key.
Method 1: Below is the syntax to load a MAP
by the first method: Use an arrow (->) to separate the map's key and value.
Method 2: The second method is to use the MAP()
function. If there are multiple key-value pairs among multiple columns, MAP()
can load them together. Below is an example:
Method 3: The third method is to use the SPLIT()
function. Similar to the SPLIT()
in loading LIST
or SET
, the SPLIT()
function can be used when the key-value pair is in one column and separated by a key-value separator, or multiple key-value pairs are in one column and separated by element separators and key-value separators. SPLIT()
here has three parameters: The first is the column index, the second is the key-value separator, and the third is the element separator. The third parameter is optional. If one row of raw data only has one key-value pair, the third parameter can be skipped. Below are the examples without and with the given element separator.
The SPLIT() function cannot be used for UDT type elements.
Loading Composite Key Attributes
Loading a Composite Key for a vertex works no differently than normal loading. Simply load all the attributes as you would for a vertex with a single-attribute primary key. The primary key will automatically be constructed from the appropriate attributes.
When loading to an edge where either TO_VERTEX
or FROM_VERTEX
contains a composite key, the composite set of attributes must be enclosed in parentheses. See the example below.
Loading Wildcard Type Edges
If an edge has been defined using a wildcard vertex type, a vertex type name must be specified, following the vertex id, in a LOAD
statement for the edge. An example is shown below:
Token Functions
Token functions are functions that operate on tokens. Some may be used to construct attribute expressions and some may be used for conditional expressions in the WHERE
clause.
To use a token function, replace the attribute value in the destination clause of the LOAD
statement with a function call.
Example
Built-in token functions for attribute expressions
The following token functions can be used in an id or attribute expression
Function | Output type | Description |
gsql_reverse( in_string ) | string | Returns a string with the characters in the reverse order of the input string in_string. |
gsql_concat( string1, string2,...,stringN ) | string | Returns a string which is the concatenation of all the input strings. |
gsql_uuid_v4() | string | Returns a version-4 UUID. |
gsql_split_by_space( in_string ) | string | Returns a modified version of in_string, in which each space character is replaced with ASCII 30 (decimal). |
gsql_substring(str, beginIndex [, length]) | string | Returns the substring beginning at beginIndex, having the given length. |
gsql_find(str, substr ) | int | Returns the start index of the substring within the string. If it is not found, then return -1. |
gsql_length(str ) | int | Returns the length of the string. |
gsql_replace(str, oldToken, newToken [, max]) | string | Returns the string resulting from replacing all matchings of oldToken with newToken in the original string. If a max count is provided, there can only be up to that many replacements. |
gsql_regex_replace( str, regex, replaceSubstr ) | string | Returns the string resulting from replacing all substrings in the input string that match the given regex token with the substitute string. |
gsql_regex_match(str, regex ) | bool | Returns true if the given string token matches the given regex token and false otherwise. |
gsql_to_bool( in_string ) | bool | Returns true if the in_string is either "t" or "true", with case insensitive checking. Returns false otherwise. |
gsql_to_uint( in_string ) | uint | If in_string is the string representation of an unsigned int, the function returns that integer. If in_string is the string representation of a nonnegative float, the function returns that number cast as an int. |
gsql_to_int( in_string ) | int | If in_string is the string representation of an int, the function returns that integer. If in_string is the string representation of a float, the function returns that number cast as an int. |
gsql_ts_to_epoch_seconds( timestamp ) | uint | Converts a timestamp in canonical string format to Unix epoch time, which is the int number of seconds since Jan. 1, 1970. Refer to the timestamp input format note below. |
gsql_current_time_epoch(0) | uint | Returns the current time in Unix epoch seconds. *By convention, the input parameter should be 0, but it is ignored. |
flatten( column_to_be_split, group_separator, 1 ) flatten( column_to_be_split, group_separator, sub_field_separator, number_of_sub_fields_in_one_group ) | See the section "TEMP_TABLE and Flatten Functions" below. | |
flatten_json_array ( $"array_name" ) flatten_json_array ( $"array_name", $"sub_obj_1", $"sub_obj_2", ..., $"sub_obj_n" ) | See the section "TEMP_TABLE and Flatten Functions" below. | |
split( column_to_be_split, element_separator ) split( column_to_be_split, key_value_separator, element _separator ) | See the section "Loading a LIST or SET Attribute" above. See the section "Loading a MAP Attribute" above. | |
gsql_upper( in_string ) | string | Returns the input string in upper-case. |
gsql_lower( in_string ) | string | Returns the input string in lower-case. |
gsql_trim( in_string ) | string | Trims whitespace from the beginning and end of the input string. |
gsql_ltrim( in_string ) gsql_rtrim( in_string ) | string | Trims white space from either the beginning or the end of the input string (Left or right). |
gsql_year( timestamp ) | int | Returns 4-digit year from timestamp. Refer to timestamp input format note below. |
gsql_month( timestamp ) | int | Returns month (1-12) from timestamp. Refer to timestamp input format note below. |
gsql_day( timestamp ) | int | Returns day (1-31) from timestamp. Refer to timestamp input format note below. |
gsql_year_epoch( epoch ) | int | Returns 4-digit year from Unix epoch time, which is the int number of seconds since Jan. 1, 1970. |
gsql_month_epoch( epoch ) | int | Returns month (1-12) from Unix epoch time, which is the int number of seconds since Jan. 1, 1970. |
gsql_day_epoch( epoch ) | int | Returns day (1-31) from Unix epoch time, which is the int number of seconds since Jan. 1, 1970. |
Timestamp Input Format
The timestamp parameter should be in one of the following formats:
"%Y-%m-%d %H:%M:%S"
"%Y/%m/%d %H:%M:%S"
"%Y-%m-%dT%H:%M:%S.000z" // text after the dot . is ignored
User-Defined Token Functions
Users can write their own token functions in C++ and install them in the GSQL system. To learn how to add a user-defined token function, see Add a User-Defined Token Function.
Reducer Functions
A reducer function aggregates multiple values of a non-id attribute into one attribute value of a single vertex or edge. Reducer functions are computed incrementally; that is, each time a new input token is applied, a new resulting value is computed.
To reduce and load aggregate data to an attribute, the attribute expression has the form
where reducer_function is one of the functions in the table below. input_expr can include non-reducer functions, but reducer functions cannot be nested.
Each reducer function is overloaded so that one function can be used for several different data types. For primitive data types, the output type is the same as the input_expr type. For LIST, SET, and MAP containers, the input_expr type is one of the allowed element types for these containers (see "Complex Types" in the Attribute Data Types section). The output is the entire container.
Function name | Data type of arg: Description of function's return value |
max( arg ) | INT, UINT, FLOAT, DOUBLE: maximum of all arg values cumulatively received |
min( arg ) | INT, UINT, FLOAT, DOUBLE: minimum of all arg values cumulatively received |
add( arg ) | INT, UINT, FLOAT, DOUBLE: sum of all arg values cumulatively received STRING: concatenation of all arg values cumulatively received LIST, SET element: list/set of all arg values cumulatively received MAP (key -> value) pair: key-value dictionary of all key-value pair arg values cumulatively received |
and( arg ) | BOOL: AND of all arg values cumulatively received INT, UINT: bitwise AND of all arg values cumulatively received |
or( arg ) | BOOL: OR of all arg values cumulatively received INT, UINT: bitwise OR of all arg values cumulatively received |
overwrite( arg ) | non-container: arg LIST, SET: new list/set containing only arg |
ignore_if_exists( arg ) | Any: If an attribute value already exists, return(retain) the existing value. Otherwise, return(load) arg . |
Each function supports a certain set of attribute types. Calling a reducer function with an incompatible type crashes the service. In order to prevent that, use the WHERE clause (introduced below) together with IS NUMERIC or other operators, functions, predicates for type checking if necessary.
WHERE Clause
The WHERE
clause is an optional clause. The WHERE
clause's condition is a boolean expression. The expression may use column token variables, token functions, and operators which are described below. The expression is evaluated for each input data line. If the condition is true, then the vertex or edge instance is loaded into the graph store. If the condition is false, then this instance is skipped. Note that all attribute values are treated as string values in the expression, so the type conversion functions to_int()
and to_float()
, which are described below, are provided to enable numerical conditions.
Operators in the WHERE Clause
The GSQL Loader language supports most of the standard arithmetic, relational, and boolean operators found in C++. Standard operator precedence applies, and parentheses provide the usual override of precedence.
Arithmetic Operators: +, -, *, /, ^ Numeric operators can be used to express complex operations between numeric types. Just as in ordinary mathematical expressions, parentheses can be used to define a group and to modify the order of precedence.
Because computers necessarily can only store approximations for most DOUBLE
and FLOAT
type values, it is not recommended to test these data types for exact equality or inequality. Instead, one should allow for an acceptable amount of error. The following example checks if $0 = 5
, with an error of 0.00001 permitted:
Relational Operators: <, >, ==, !=, <=, >= Comparisons can be performed between two numeric values or between two string values.
Predicate Operators:
AND, OR, NOT operators are the same as in SQL. They can be used to combine multiple conditions together. E.g., $0 < "abc" AND $1 > "abc" selects the rows with the first token less than "abc" and the second token greater than "abc". E.g., NOT $1 < "abc" selects the rows with the second token greater than or equal to "abc".
IS NUMERIC token IS NUMERIC returns true if token is in numeric format. Numeric format include integers, decimal notation, and exponential notation. Specifically, IS NUMERIC is true if token matches the following regular expression: (+/-) ? [0-9] + (.[0-9]) ? [0-9] * ((e/E)(+/-) ? [0-9] +) ? . Any leading space and trailing space is skipped, but no other spaces are allowed. E.g., $0 IS NUMERIC checks whether the first token is in numeric format.
IS EMPTY token IS EMPTY returns true if token is an empty string. E.g., $1 IS EMPTY checks whether the second token is empty.
IN token IN ( set_of_values ) returns true if token is equal to one member of a set of specified values. The values may be string or numeric types. E.g., $2 IN ("abc", "def", "lhm") tests whether the third token equals one of the three strings in the given set. E.g., to_int($3) IN (10, 1, 12, 13, 19) tests whether the fourth token equals one of the specified five numbers.
BETWEEN ... AND token BETWEEN lowerVal AND upperVal returns true if token is within the specified range, inclusive of the endpoints. The values may be string or numeric types. E.g., $4 BETWEEN "abc" AND "def" checks whether the fifth token is greater than or equal to "abc" and also less than or equal to "def" E.g., to_float($5) BETWEEN 1 AND 100.5 checks whether the sixth token is greater than or equal to 1.0 and less than or equal to 100.5.
Token functions in the WHERE clause
The GSQL loading language provides several built-in functions for the WHERE clause.
Function name | Output type | Description of function |
to_int( main_string ) | int | Converts main_string to an integer value. |
to_float( main_string ) | float | Converts main_string to a float value. |
concat( string1, string2 ) | string | Returns a string which is the concatenation of string1 and string2 . |
token_len( main_string ) | int | Returns the length of main_string. |
gsql_is_not_empty_string( main_string ) | bool | Returns true if main_string is empty after removing white space. Returns false otherwise. |
gsql_token_equal( string1, string2 ) | bool | Returns true if string1 is exactly the same (case sensitive) as string2 . Returns false otherwise. |
gsql_token_ignore_case_equal( string1, string2 ) | bool | Returns true if string1 is exactly the same (case insensitive) as string2 . Returns false otherwise. |
gsql_is_true( main_string ) | bool | Returns true if main_string is either "t" or "true" (case insensitive). Returns false otherwise. |
gsql_is_false( main_string ) | bool | Returns true if main_string is either "f" or "false" (case insensitive). Returns false otherwise. |
The token functions in the WHERE clause and those token functions used for attribute expression are different. They cannot be used exchangeably.
Other Optional LOAD Clauses
OPTION
clause
OPTION
clauseThere are no supported options for the OPTION clause at this time.
TAGS
clause (Beta)
TAGS
clause (Beta)The TAGS
clause specifies the tags to be applied to the vertices loaded by the LOAD
statement.
If a LOAD
statement has a TAGS
clause, it will tag the vertices with the tags specified in the TAGS
clause. Before vertices can be loaded and tagged with a LOAD
statement, the vertex type must first be marked as taggable, and the tags must be defined.
Users have two options when it comes to how to merge tags if the target vertices exist in the graph:
BY OR
: Add the new tags to the existing set of tags.BY OVERWRITE
: Overwrite existing tags with the new tags.
USING
clause
USING
clauseA USING
clause contains one or more optional parameter value pairs:
If multiple LOAD statements use the same source (the same file path, the same TEMP_TABLE, or the same file variable), the USING clauses in these LOAD statements must be the same. Therefore, we recommend that if multiple destination clauses share the same source, put all of these destination clauses into the same LOAD statement.
Parameter | Meaning of Value | Allowed Values |
SEPARATOR | specifies the special character that separates tokens (columns) in the data file | any single ASCII character. Default is comma ","
|
EOL | the end-of-line character | any ASCII sequence Default = |
QUOTE (See note below) | specifies explicit boundary markers for string tokens, either single or double quotation marks. See more details below. | "single" for ' "double" for " |
HEADER | whether the data file's first line is a header line. The header assigns names to the columns. The LOAD statement must refer to an actual file with a valid header. | "true", "false" Default is "false" |
USER_DEFINED_HEADER | specifies the name of the header variable, when a header has been defined in the loading job, rather than in the data file | the variable name in the preceding DEFINE HEADER statement |
REJECT_LINE_RULE | if the filter expression evaluates to true, then do not use this input data line. | name of filter from a preceding DEFINE INPUT_LINE_FILTER statement |
JSON_FILE (See Loading JSON Data section below) | whether each line is a json object (see Section "JSON Loader" below for more details) | "true", "false" Default is "false" |
NEW_VERTEX_ONLY | If true, treat vertices as insert-only. If the input data refers to a vertex which already exists, do not update it. If false, upsert vertices. | "true", "false" Default is "false" |
VERTEX_MUST_EXIST (See VERTEX_MUST_EXIST section below) | If true, only insert or update an edge If both endpoint vertices already exist. If false, always insert new edges, creating endpoint vertices as needed, using given id and default values for other parameters. | "true", "false" Default is "false" |
QUOTE parameter
The parser will not treat separator characters found within a pair of quotation marks as a separator. For example, if the parsing conditions are QUOTE="double", SEPARATOR=","
, the comma in "Leonard,Euler"
will not separate Leonard and Euler into separate tokens.
If QUOTE is not declared, quotation marks are treated as ordinary characters.
If QUOTE is declared, but a string does not contain a matching pair of quotation marks, then the string is treated as if QUOTE is not declared.
Only the string inside the first pair of quote (from left to right) marks are loaded. For example QUOTE="double", the string a"b"c"d"e will be loaded as b.
There is no escape character in the loader, so the only way to include quotation marks within a string is for the string body to use one type of quote (single or double) and to declare the other type as the string boundary marker.
Previously, ill-formatted strings such as a"a,b"ac,d would be parsed as a,b,d ignoring a,a,c. The expected input string should be a,"a,b",ac,d. In v2.4, incorrectly formatted strings such as this example will be parsed normally, giving you this result: a"a,b"ac and d.
Loading JSON Data
When the USING option JSON_FILE="true"
is used, the loader loads JSON objects instead of tabular data. A JSON object is an unordered set of key/value pairs, where each value may itself be an array or object, leading to nested structures. A colon separates each key from its value, and a comma separates items in a collection. A more complete description of JSON format is available at www.json.org. The JSON loader requires that each input line has exactly one JSON object. Instead of using column values as tokens, the JSON loader uses JSON values as tokens, that is, the second part of each JSON key/value pair. In a GSQL loading job, a JSON field is identified by a dollar sign $ followed by the colon-separated sequence of nested key names to reach the value from the top level. For example, given the JSON object {"abc":{"def": "this_value"}}
, the identifier $"abc":"def"
is used to access "this_value"
. The double quotes are mandatory.
An example is shown below:
To specify an end-of-line character other than the standard one, use the EOL option, as shown below.
In the above data encoding.json, the order of fields are not fixed and some fields are missing. The JSON loader ignores the order and accesses the fields by the nested key names. The missing fields are loaded with default values. The result vertices are:
id | attr1 |
"UTF-7" | 30 |
"UTF-1" | 0 |
"UTF-6" | 3 |
Loading Parquet Data
TigerGraph can load data from Parquet files if they are stored in AWS S3 buckets. For more details on how to set up S3 data sources and loading jobs, read the AWS S3 Loader User Guide. In the background TigerGraph uses the JSON loading functionality to read data from Parquet files, so the JSON specific information in the previous section applies.
In order to load Parquet data, you need to:
Specify
"file.reader.type": "parquet"
in the S3 file configuration file or argumentSpecify
JSON_FILE="true"
in the USING clause of the LOAD statementsRefer to JSON keys (≈ Parquet "column names") instead of column numbers
You will probably want to add USING EOF="true"
to your RUN LOADING JOB statement to explicitly indicate to the loading job to stop after consuming all data from the Parquet source, not to expect further entries.
An example of a Parquet loading setup is shown below:
VERTEX_MUST_EXIST
Parameter
VERTEX_MUST_EXIST
ParameterNormally, if vertices do not exist when loading data to edges, a vertex will be created for the connecting edge, using default values for all attributes. Using the VERTEX_MUST_EXIST="true" option will load data only if the vertices on both sides of an edge already exist, therefore no longer creating extra vertices.
TEMP_TABLE
and Flatten Functions
TEMP_TABLE
and Flatten FunctionsThe keyword TEMP_TABLE
triggers the use of a temporary data table which is used to store data generated by one LOAD
statement, for use by a later LOAD
statement. Earlier we introduced the syntax for loading data to a TEMP_TABLE
:
This clause is designed to be used in conjunction with the flatten
or flatten_json_array
function in one of the attr_expr
expressions. The flatten function splits a multi-value field into a set of records. Those records can first be stored in a temporary table, and then the temporary table can be loaded into vertices and/or edges. Only one flatten function is allowed in one temp table destination clause.
There are two versions of the flatten function: One parses single-level groups and the other parses two-level groups. There are also two versions of the flatten_json_array function: One splits an array of primitive values, and the other splits an array of JSON objects.
One-Level Flatten Function
flatten( column_to_be_split, separator, 1 )
is used to parse a one-level group into individual elements. An example is shown below:
The following loading job contains two LOAD
statements. The first one loads input data to Book
vertices and to a TEMP_TABLE
. The second one loads the TEMP_TABLE
data to Genre
vertices and book_genre
edges.
Line 5 says that the third column ($2
) of each input line should be split into separate tokens, with comma (,
) as the separator. Each token will have its own row in table t1
. The first column is labeled bookcode
with value $0
and the second column is genre
with one of the $2
tokens. The contents of TEMP_TABLE t1
are shown below:
bookcode | genre |
101 | fiction |
101 | fantasy |
101 | young_adult |
102 | fiction |
102 | science_fiction |
102 | Chinese |
Then, lines 8 to 10 say to read TEMP_TABLE t1
and to do the following for each row:
Create a
Genre
vertex for each new value ofgenre
.Create a
book_genre
edge frombookcode
togenre
. In this case, each row ofTEMP_TABLE t1
generates onebook_genre
edge.
The final graph will contain two Book
vertices (101 and 102), five Genre vertices, and six book_genre
edges.
Two-Level Flatten Function
flatten( column_to_be_split, group_separator, sub_field_separator, number_of_sub_fields_in_one_group ) is used for parse a two-level group into individual elements. Each token in the main group may itself be a group, so there are two separators: one for the top level and one for the second level. An example is shown below.
The flatten function now has four parameters instead of three. The additional parameter is used to record the genre_name in the Genre vertices.
In this example, in the genres column ($2), there are multiple groups, and each group has two sub-fields, genre_id and genre_name. After running the loading job, the file book2.dat will be loaded into the TEMP_TABLE t2 as shown below.
bookcode | genre_id | |
101 | FIC | fiction |
101 | FTS | fantasy |
101 | YA | young adult |
102 | FIC | fiction |
102 | SF | science fiction |
102 | CHN | Chinese |
Flatten a JSON Array of Primitive Values
flatten_json_array($" array_name ") parses a JSON array of primitive (string, numberic, or bool) values, where "array_name" is the name of the array. Each value in the array creates a record. Below is an example:
The above data and loading job creates the following temporary table:
id | length |
C | 3 |
c++ | 3 |
Flatten a JSON Array of JSON Objects
flatten_json_array ( $"array_name", $"sub_obj_1", $"sub_obj_2", ..., $"sub_obj_n" ) parses a JSON array of JSON objects. "array_name" is the name of the array, and the following parameters $"sub_obj_1", $"sub_obj_2", ..., $"sub_obj_n" are the field key names in each object in the array. See complete example below:
When splitting a JSON array of JSON objects, the primitive values are skipped and only JSON objects are processed. As in the example above, the 4th line's "plug-ins" field will not generate any record because its "plug-ins" array doesn't contain any JSON object. Any field which does not exist in the object will be loaded with default value. The above example generates the temporary table shown below:
id | score | age | length |
"golang" | default | "noidea" | default |
"pascal" | 1.0 | "old" | 12 |
"c++" | 2.0 | default | 12 |
"java" | 2.22 | default | 30 |
"python" | 3.0 | default | 30 |
"go" | 4.0 | "new" | 30 |
Flatten a JSON Array in a CSV file
flatten_json_array() can also be used to split a column of a tabular file, where the column contains JSON arrays. An example is given below:
The second column in the csv file is a JSON array which we want to split. flatten_json_array() can be used in this case without the USING JSON_FILE="true" clause:
The above example generates the temporary table shown below:
id | score | age | length |
golang | -1 (default) | noidea | -1 (default) |
pascal | 1 | old | -1 (default) |
c++ | 2 | unknown (default) | 12 |
java | 2.22 | new | 2 |
python | 3 | unknown (default) | 4 |
go | 4 | new | -1 (default) |
flatten_json_array in csv
flatten_json_array() does not work if the separator appears also within the json array column. For example, if the separator is comma, the csv loader will erroneously divide the json array into multiple columns. Therefore, it is recommended that the csv file use a special column separator, such as "|" in the above example .
DELETE statement
In addition to loading data, a LOADING JOB can be used to perform the opposite operation: deleting vertices and edges, using the DELETE statement. DELETE cannot be used in offline loading. Just as a LOAD statement uses the tokens from each input line to set the id and attribute values of a vertex or edge to be created, a DELETE statement uses the tokens from each input line to specify the id value of the item(s) to be deleted.
In the v2.0 syntax, there is now a " FROM (filepath_string | filevar)
" clause just before the WHERE clause.
There are four variations of the DELETE statement. The syntax of the four cases is shown below.
An example using book_rating data is shown below:
There is a separate DELETE statement in the GSQL Query Language. The query delete statement can leverage the query language's ability to explore the graph and to use complex conditions to determine which items to delete. In contrast, the loading job delete statement requires that the id values of the items to be deleted must be specified in advance in an input file.
Last updated