Extending

UDFs (User Defined Functions)

Our expression engine can be extended with user defined functions, or UDFs for short, like this:

SELECT id, attr1, myudf(attr2, attr3+attr4) ...

You can load and unload UDFs dynamically into searchd without having to restart the daemon, and used them in expressions when searching, ranking, etc. Quick summary of the UDF features is as follows.

  • UDFs can take integer (both 32-bit and 64-bit), float, string, MVA, or PACKEDFACTORS() arguments.
  • UDFs can return integer, float, or string values.
  • UDFs can check the argument number, types, and names during the query setup phase, and raise errors.
  • Aggregation UDFs are not yet supported (but might be in the future).

UDFs have a wide variety of uses, for instance:

  • adding custom mathematical or string functions;
  • accessing the database or files from within Manticore;
  • implementing complex ranking functions.

UDFs reside in the external dynamic libraries (.so files on UNIX and .dll on Windows systems). Library files need to reside in a trusted folder specified by plugin_dir directive, for obvious security reasons: securing a single folder is easy; letting anyone install arbitrary code into searchd is a risk. You can load and unload them dynamically into searchd with CREATE FUNCTION and DROP FUNCTION SphinxQL statements respectively. Also, you can seamlessly reload UDFs (and other plugins) with RELOAD PLUGINS statement. Manticore keeps track of the currently loaded functions, that is, every time you create or drop an UDF, searchd writes its state to the sphinxql_state file as a plain good old SQL script.

Once you successfully load an UDF, you can use it in your SELECT or other statements just as well as any of the builtin functions:

SELECT id, MYCUSTOMFUNC(groupid, authorname), ... FROM myindex

Multiple UDFs (and other plugins) may reside in a single library. That library will only be loaded once. It gets automatically unloaded once all the UDFs and plugins from it are dropped.

In theory you can write an UDF in any language as long as its compiler is able to import standard C header, and emit standard dynamic libraries with properly exported functions. Of course, the path of least resistance is to write in either C++ or plain C. We provide an example UDF library written in plain C and implementing several functions (demonstrating a few different techniques) along with our source code, see src/udfexample.c. That example includes src/sphinxudf.h header file definitions of a few UDF related structures and types. For most UDFs and plugins, a mere #include "sphinxudf.h", like in the example, should be completely sufficient, too. However, if you’re writing a ranking function and need to access the ranking signals (factors) data from within the UDF, you will also need to compile and link with src/sphinxudf.c (also available in our source code), because the implementations of the fuctions that let you access the signal data from within the UDF reside in that file.

Both sphinxudf.h header and sphinxudf.c are standalone. So you can copy around those files only; they do not depend on any other bits of Manticore source code.

Within your UDF, you must implement and export only a couple functions, literally. First, for UDF interface version control, you must define a function int LIBRARYNAME_ver(), where LIBRARYNAME is the name of your library file, and you must return SPH_UDF_VERSION (a value defined in sphinxudf.h) from it. Here’s an example.

#include <sphinxudf.h>

// our library will be called udfexample.so, thus, so it must define
// a version function named udfexample_ver()
int udfexample_ver()
{
    return SPH_UDF_VERSION;
}

That protects you from accidentally loading a library with a mismatching UDF interface version into a newer or older searchd. Second, yout must implement the actual function, too. sphinx_int64_t testfunc ( SPH_UDF_INIT * init, SPH_UDF_ARGS * args, char * error_flag ) { return 123; }

UDF function names in SphinxQL are case insensitive. However, the respective C function names are not, they need to be all lower-case, or the UDF will not load. More importantly, it is vital that a) the calling convention is C (aka __cdecl), b) arguments list matches the plugin system expectations exactly, and c) the return type matches the one you specify in CREATE FUNCTION. Unfortunately, there is no (easy) way for us to check for those mistakes when loading the function, and they could crash the server and/or result in unexpected results. Last but not least, all the C functions you implement need to be thread-safe.

The first argument, a pointer to SPH_UDF_INIT structure, is essentially a pointer to our function state. It is option. In the example just above the function is stateless, it simply returns 123 every time it gets called. So we do not have to define an initialization function, and we can simply ignore that argument.

The second argument, a pointer to SPH_UDF_ARGS, is the most important one. All the actual call arguments are passed to your UDF via this structure; it contians the call argument count, names, types, etc. So whether your function gets called like SELECT id, testfunc(1) or like SELECT id, testfunc('abc', 1000*id+gid, WEIGHT()) or anyhow else, it will receive the very same SPH_UDF_ARGS structure in all of these cases. However, the data passed in the args structure will be different. In the first example args->arg_count will be set to 1, in the second example it will be set to 3, args->arg_types array will contain different type data, and so on.

Finally, the third argument is an error flag. UDF can raise it to indicate that some kinda of an internal error happened, the UDF can not continue, and the query should terminate early. You should not use this for argument type checks or for any other error reporting that is likely to happen during normal use. This flag is designed to report sudden critical runtime errors, such as running out of memory.

If we wanted to, say, allocate temporary storage for our function to use, or check upfront whether the arguments are of the supported types, then we would need to add two more functions, with UDF initialization and deinitialization, respectively.

int testfunc_init ( SPH_UDF_INIT * init, SPH_UDF_ARGS * args,
    char * error_message )
{
    // allocate and initialize a little bit of temporary storage
    init->func_data = malloc ( sizeof(int) );
    *(int*)init->func_data = 123;

    // return a success code
    return 0;
}

void testfunc_deinit ( SPH_UDF_INIT * init )
{
    // free up our temporary storage
    free ( init->func_data );
}

Note how testfunc_init() also receives the call arguments structure. By the time it is called it does not receive any actual values, so the args->arg_values will be NULL. But the argument names and types are known and will be passed. You can check them in the initialization function and return an error if they are of an unsupported type.

UDFs can receive arguments of pretty much any valid internal Manticore type. Refer to sphinx_udf_argtype enumeration in sphinxudf.h for a full list. Most of the types map straightforwardly to the respective C types. The most notable exception is the SPH_UDF_TYPE_FACTORS argument type. You get that type by calling your UDF with a PACKEDFACTOR() argument. It’s data is a binary blob in a certain internal format, and to extract individual ranking signals from that blob, you need to use either of the two sphinx_factors_XXX() or sphinx_get_YYY_factor() families of functions. The first family consists of just 3 functions, sphinx_factors_init() that initializes the unpacked SPH_UDF_FACTORS structure, sphinx_factors_unpack() that unpacks a binary blob into it, and sphinx_factors_deinit() that cleans up an deallocates the SPH_UDF_FACTORS. So you need to call init() and unpack(), then you can use the SPH_UDF_FACTORS fields, and then you need to cleanup with deinit(). That is simple, but results in a bunch of memory allocations per each processed document, and might be slow. The other interface, consisting of a bunch of sphinx_get_YYY_factor() functions, is a little more wordy to use, but accesses the blob data directly and guarantees that there will be zero allocations. So for top-notch ranking UDF performance, you want to use that one.

As for the return types, UDFs can currently return a signle INTEGER, BIGINT, FLOAT, or STRING value. The C function return type should be sphinx_int64_t, sphinx_int64_t, double, or char* respectively. In the last case you must use args->fn_malloc function to allocate the returned string values. Internally in your UDF you can use whatever you want, so the testfunc_init() example above is correct code even though it uses malloc() directly: you manage that pointer yourself, it gets freed up using a matching free() call, and all is well. However, the returned strings values are managed by Manticore and we have our own allocator, so for the return values specifically, you need to use it too.

Depending on how your UDFs are used in the query, the main function call (testfunc() in our example) might be called in a rather different volume and order. Specifically,

  • UDFs referenced in WHERE, ORDER BY, or GROUP BY clauses must and will be evaluated for every matched document. They will be called in the natural matching order.
  • without subselects, UDFs that can be evaluated at the very last stage over the final result set will be evaluated that way, but before applying the LIMIT clause. They will be called in the result set order.
  • with subselects, such UDFs will also be evaluated after applying the inner LIMIT clause.

The calling sequence of the other functions is fixed, though. Namely,

  • testfunc_init() is called once when initializing the query. It can return a non-zero code to indicate a failure; in that case query will be terminated, and the error message from the error_message buffer will be returned.
  • testfunc() is called for every eligible row (see above), whenever Manticore needs to compute the UDF value. It can also indicate an (internal) failure error by writing a non-zero byte value to error_flag. In that case, it is guaranteed that will no more be called for subsequent rows, and a default return value of 0 will be substituted. Manticore might or might not choose to terminate such queries early, neither behavior is currently guaranteed.
  • testfunc_deinit() is called once when the query processing (in a given index shard) ends.

We do not yet support aggregation functions. In other words, your UDFs will be called for just a single document at a time and are expected to return some value for that document. Writing a function that can compute an aggregate value like AVG() over the entire group of documents that share the same GROUP BY key is not yet possible. However, you can use UDFs within the builtin aggregate functions: that is, even though MYCUSTOMAVG() is not supported yet, AVG(MYCUSTOMFUNC()) should work alright!

UDFs are local. In order to use them on a cluster, you have to put the same library on all its nodes and run CREATEs on all the nodes too. This might change in the future versions.

Plugins

Here’s the complete plugin type list.

  • UDF plugins;
  • ranker plugins;
  • indexing-time token filter plugins;
  • query-time token filter plugins.

This section discusses writing and managing plugins in general; things specific to writing this or that type of a plugin are then discussed in their respective subsections.

So, how do you write and use a plugin? Four-line crash course goes as follows:

  • create a dynamic library (either .so or.dll), most likely in C or C++;
  • load that plugin into searchd using CREATE PLUGIN;
  • invoke it using the plugin specific calls (typically using this or that OPTION).
  • to unload or reload a plugin use DROP PLUGIN and RELOAD PLUGINS respectively.

Note that while UDFs are first-class plugins they are nevertheless installed using a separate CREATE FUNCTION statement. It lets you specify the return type neatly so there was especially little reason to ruin backwards compatibility and change the syntax.

Dynamic plugins are supported in threads and thread_pool workers. Multiple plugins (and/or UDFs) may reside in a single library file. So you might choose to either put all your project-specific plugins in a single common uber-library; or you might choose to have a separate library for every UDF and plugin; that is up to you.

Just as with UDFs, you want to include src/sphinxudf.h header file. At the very least, you will need the SPH_UDF_VERSION constant to implement a proper version function. Depending on the specific plugin type, you might or might not need to link your plugin with src/sphinxudf.c. However, all the functions implemented in sphinxudf.c are about unpacking the PACKEDFACTORS() blob, and no plugin types are exposed to that kind of data. So currently, you would never need to link with the C-file, just the header would be sufficient. (In fact, if you copy over the UDF version number, then for some of the plugin types you would not even need the header file.)

Formally, plugins are just sets of C functions that follow a certain naming parttern. You are typically required to define just one key function that does the most important work, but you may define a bunch of other functions, too. For example, to implement a ranker called “myrank”, you must define myrank_finalize() function that actually returns the rank value, however, you might also define myrank_init(), myrank_update(), and myrank_deinit() functions. Specific sets of well-known suffixes and the call arguments do differ based on the plugin type, but _init() and _deinit() are generic, every plugin has those. Protip: for a quick reference on the known suffixes and their argument types, refer to sphinxplugin.h, we define the call prototoypes in the very beginning of that file.

Despite having the public interface defined in ye good olde good pure C, our plugins essentially follow the object-oriented model. Indeed, every _init() function receives a void ** userdata out-parameter. And the pointer value that you store at (*userdata) location is then be passed as a 1st argument to all the other plugin functions. So you can think of a plugin as class that gets instantiated every time an object of that class is needed to handle a request: the userdata pointer would be its this pointer; the functions would be its methods, and the _init() and _deinit() functions would be the constructor and destructor respectively.

Why this (minor) OOP-in-C complication? Well, plugins run in a multi-threaded environment, and some of them have to be stateful. You can’t keep that state in a global variable in your plugin. So we have to pass around a userdata parameter anyway to let you keep that state. And that naturally brings us to the OOP model. And if you’ve got a simple, stateless plugin, the interface lets you omit the _init() and _deinit() and whatever other functions just as well.

To summarize, here goes the simplest complete ranker plugin, in just 3 lines of C code.

// gcc -fPIC -shared -o myrank.so myrank.c
#include "sphinxudf.h"
int myrank_ver() { return SPH_UDF_VERSION; }
int myrank_finalize(void *u, int w) { return 123; }

And this is how you use it:

mysql> CREATE PLUGIN myrank TYPE 'ranker' SONAME 'myrank.dll';
Query OK, 0 rows affected (0.00 sec)

mysql> SELECT id, weight() FROM test1 WHERE MATCH('test')
    -> OPTION ranker=myrank('');
+------+----------+
| id   | weight() |
+------+----------+
|    1 |      123 |
|    2 |      123 |
+------+----------+
2 rows in set (0.01 sec)

Ranker plugins

Ranker plugins let you implement a custom ranker that receives all the occurrences of the keywords matched in the document, and computes a WEIGHT() value. They can be called as follows:

SELECT id, attr1 FROM test WHERE match('hello')
OPTION ranker=myranker('option1=1');

The call workflow is as follows:

  1. XXX_init() gets called once per query per index, in the very beginning. A few query-wide options are passed to it through a SPH_RANKER_INIT structure, including the user options strings (in the example just above, “option1=1” is that string).
  2. XXX_update() gets called multiple times per matched document, with every matched keyword occurrence passed as its parameter, a SPH_RANKER_HIT structure. The occurrences within each document are guaranteed to be passed in the order of ascending hit->hit_pos values.
  3. XXX_finalize() gets called once per matched document, once there are no more keyword occurrences. It must return the WEIGHT() value. This is the only mandatory function.
  4. XXX_deinit() gets called once per query, in the very end.

Token filter plugins

Token filter plugins let you implement a custom tokenizer that makes tokens according to custom rules. There are two type:

Token filters processing tokens after base tokenizer processed text at field or query and made tokens from it. In the text processing pipeline, the token filters will run after the base tokenizer processing occurs (which process the text from field or query and create tokens out of them).

Index-time tokeniker

Index-time tokenizer gets created by indexer on indexing source data into index or by RT index on processing INSERT or REPLACE statements.

Plugin is declared as library name:plugin name:optional string of settings. The init functions of the plugin can accept arbitrary settings that can be passed as a string in format option1=value1;option2=value2;...

Example:

index_token_filter = my_lib.so:email_process:field=email;split=.io

The call workflow for index-time token filter is as follows:

  1. XXX_init() gets called right after indexer creates token filter with empty fields list then after indexer got index schema with actual fields list. It must return zero for successful initialization or error description otherwise.

  2. XXX_begin_document gets called only for RT index INSERT/REPLACE for every document. It must return zero for successful call or error description otherwise. Using OPTION token_filter_options additional parameters/settings can be passed to the function.

    INSERT INTO rt (id, title) VALUES (1, 'some text corp@space.io') OPTION token_filter_options='.io'
    
  3. XXX_begin_field gets called once for each field prior to processing field with base tokenizer with field number as its parameter.

  4. XXX_push_token gets called once for each new token produced by base tokenizer with source token as its parameter. It must return token, count of extra tokens made by token filter and delta position for token.

  5. XXX_get_extra_token gets called multiple times in case XXX_push_token reports extra tokens. It must return token and delta position for that extra token.

  6. XXX_end_field gets called once right after source tokens from current field get over.

  7. XXX_deinit gets called in the very end of indexing.

The following functions are mandatory to be defined: XXX_begin_document and XXX_push_token and XXX_get_extra_token.

query-time token filter

Query-time tokenizer gets created on search each time full-text invoked by every index involved.

The call workflow for query-time token filter is as follows:

  1. XXX_init() gets called once per index prior to parsing query with parameters - max token length and string set by token_filter option

    SELECT * FROM index WHERE MATCH ('test') OPTION token_filter='my_lib.so:query_email_process:io'
    

    It must return zero for successful initialization or error description otherwise.

  2. XXX_push_token() gets called once for each new token produced by base tokenizer with parameters: token produced by base tokenizer, pointer to raw token at source query string and raw token length. It must return token and delta position for token.

  3. XXX_pre_morph() gets called once for token right before it got passed to morphology processor with reference to token and stopword flag. It might set stopword flag to mark token as stopword.

  4. XXX_post_morph() gets called once for token after it processed by morphology processor with reference to token and stopword flag. It might set stopword flag to mark token as stopword. It must return flag non-zero value of which means to use token prior to morphology processing.

  5. XXX_deinit() gets called in the very end of query processing.

Absence of any of the functions is tolerated.