UDFs need to be threadsafe with pseudo_sharding ?

General Discussions - English
1

I have have some basic UDFs that keep track of results to help create some various shuffling etc.

Someone has just reported the longer work with pseudo_sharding (enabled by default in recent versions). Manticore crashes.

The plugins work by setting up a std::map in ‘udf_init’ and deallocating it in ‘udf_deinit’.

Very little detail on how pseudo_sharding works, but assuming it doing some sort of multi-threading.

Is there any special considerations UDFs need to take in light of new pseudo_sharding ? Could it be that _init and _deinit are not getting called exactly once per query?

The actual update the code taht runs, does a single ++var so in theory shouldbe atomic?

DLLEXPORT sphinx_int64_t uniqueserial(SPH_UDF_INIT *init, SPH_UDF_ARGS *args, char *)
{
    unsigned int unique = *(unsigned int*)args->arg_values[0];
    IntCount *m = static_cast<IntCount*>(init->func_data);
    return ++(*m)[unique];
}

But at worse it would just miscount, if multiple calls made to the main udf function, even from multiple threads? Shouldnt crash?

2

Is there any special considerations UDFs need to take in light of new pseudo_sharding ? Could it be that _init and _deinit are not getting called exactly once per query?

The actual update the code taht runs, does a single ++var so in theory shouldbe atomic?

I’ve asked the author of the pseudo sharding to comment on this.

3

No, you don’t need to change your code to work with pseudo_sharding. _init and _deinit are still called once per query, but multiple queries are run instead of just one. So each query has its unique func_data.

4

Ok thanks. That breaks the logical function of the UDFs, the point of the UDF is to ‘count’ results in the query, so intended to only have one per query.

… but that would just miscount. Doesn’t explain why it crashing thou.

Seems we will need to get a minimal reproducible example, to debug this.

5

Actually you can think of pseudo_sharding as running a query against an rt index that has several identical disk chunks. They work basically the same. When disk chunks are queried (in an rt index), each thread creates a clone of all expressions (including UDFs).

However, after having a look at the code, it seems that UDFs are a special case. When UDF expressions are cloned, __init is not called (and __deinit is called for each cloned copy).

As __init is not called for each thread, no unique func_data is provided for each thread, so the udf function can really be called from several threads.

his is an issue that we need to discuss and fix.