docs/cql/wasm.rst
This document describes the details of Wasm language support in user-defined functions (UDF). The language wasm is one of the possible languages to use, besides Lua, to implement these functions. To learn more about User-defined functions in ScyllaDB, click :ref:here <udfs>.
.. note:: Until ScyllaDB 5.2, the Wasm language was called xwasm. This name is replaced with wasm in ScyllaDB 5.4.
ScyllaDB accepts UDF's source code in WebAssembly Text ("WAT") format. The source can use and define whatever's needed for execution, including multiple helper functions and symbols. The requirements for it to be accepted as correct UDF source are that the WebAssembly module export a symbol with the same name as the function, this symbol's type should be indeed a function with correct signature, and the module export a _scylla_abi global and all symbols related to the selected :ref:ABI version <abi-versions>.
UDF's source code can be, naturally, simply coded by hand in WAT. It is not often very convenient to program directly in assembly, so here are a few tips.
Compiling to Wasm ^^^^^^^^^^^^^^^^^
Rust ....
The main supported language for Wasm UDFs is Rust. To generate WebAssembly from Rust, it's best to use the scylla-udf <https://github.com/scylladb/scylla-rust-udf>_ Rust helper library, and follow the instructions presented there.
As a short example, here's a sample Rust code which can be compiled to WebAssembly:
.. code-block:: rust
use scylla_udf::export_udf;
#[export_udf]
fn flatten(list: Vec<Vec<i32>>) -> Vec<i32> {
list.into_iter().flatten().collect()
}
The compilation instructions are described at https://github.com/scylladb/scylla-rust-udf but the commands will generally be:
.. code-block:: bash
cargo build --target=wasm32-wasip1
wasm2wat target/wasm32-wasip1/debug/flatten.wasm > flatten.wat
C .
Clang is capable of compiling C source code to Wasm and it also supports useful built-ins
for using Wasm-specific interfaces, like __builtin_wasm_memory_size and __builtin_wasm_memory_grow
for memory management.
However, there is no C helper library yet, so implementing UDFs in C is in general much more difficult than in Rust. Just to implement the fib() function that returns a special value on a NULL, you need something like this:
.. code-block:: c
const int _scylla_abi = 2;
const int SPECIAL_VALUE = 42;
void* _scylla_malloc(int size) {
return malloc(size);
}
void _scylla_free(void* ptr) {
free(ptr);
}
static long long ntohll(long long val) {
val = ((val << 8) & 0xFF00FF00FF00FF00ULL ) | ((val >> 8) & 0x00FF00FF00FF00FFULL );
val = ((val << 16) & 0xFFFF0000FFFF0000ULL ) | ((val >> 16) & 0x0000FFFF0000FFFFULL );
return (val << 32) | ((val >> 32) & 0xFFFFFFFFULL);
}
static long long htonll(long long val) {
return ntohll(val);
}
static long long fib_aux(long long n) {
if (n < 2) {
return n;
}
return fib_aux(n-1) + fib_aux(n-2);
}
long long fib(long long p) {
int size = p >> 32;
long long* p_val = (long long*)(p & 0xffffffff);
// Initialize memory for the return value
long long* ret_val = _scylla_malloc(sizeof(long long));
if (size == -1) {
*ret_val = htonll(SPECIAL_VALUE);
} else {
*ret_val = htonll(fib_aux(ntohll(*p_val)));
}
_scylla_free(p_val);
// 8 is the size of a bigint
return (long long)(8ll << 32) | (long long)ret_val;
}
// using wasi in c/c++ requires adding a main function to the program
int main(){}
And compile it with:
.. code-block:: bash
/path/to/wasm/supporting/c/compiler --sysroot=/path/to/wasi/sysroot -O2 --target=wasm32-wasip1 -Wl,--export=fib -Wl,--export=_scylla_abi -Wl,--export=_scylla_malloc -Wl,--export=_scylla_free -Wl,--no-entry fibnull.c -o fibnull.wasm
wasm2wat fibnull.wasm > fibnull.wat
The example above is particularly complicated, because it handles NULL values, which causes even integers to be serialized. Because the UDF only takes Wasm-compatible types (ints/doubles) as parameters and return values,
if we specify that the UDF RETURNS NULL ON NULL INPUT, all serialization can be avoided, and the code can be simplified to:
.. code-block:: c
const int _scylla_abi = 2;
void* _scylla_malloc(int size) {
return malloc(size);
}
void _scylla_free(void* ptr) {
free(ptr);
}
long long fib(int n) {
if (n < 2) {
return n;
}
return fib(n-1) + fib(n-2);
}
int main(){}
Because we don't need to serialize anything, the _scylla_malloc and _scylla_free methods don't need to be exported, and _scylla_abi can be set to 1, resulting in an even shorter code:
.. code-block:: c
const int _scylla_abi = 1;
long long fib(int n) {
if (n < 2) {
return n;
}
return fib(n-1) + fib(n-2);
}
int main(){}
Compilation instructions:
.. code-block:: bash
clang -O2 --target=wasm32 --no-standard-libraries -Wl,--export=fib -Wl,--export=_scylla_abi -Wl,--no-entry fib.c -o fib.wasm
wasm2wat fib.wasm > fib.wat
.. The compiled example can be viewed at test/cqlpy/test_wasm.py::test_docs_c
AssemblyScript ..............
AssemblyScript is a TypeScript-like language that compiles to WebAsembly.
Install via npm:
.. code-block:: bash
npm install -g assemblyscript
Example source code:
.. code-block:: typescript
export const _scylla_abi = [1]
export function fib(n: i32): i32 {
if (n < 2) {
return n
}
return fib(n - 1) + fib(n - 2)
}
Compile directly to WebAssembly Text format with:
.. code-block:: bash
asc fib.ts --textFile fib.wat --optimize
.. The compiled example can be viewed at test/cqlpy/test_wasm.py::test_docs_assemblyscript
Similarly to C, the AssemblyScript can only be conveniently used with RETURNS NULL ON NULL INPUT UDFs that only have Wasm-compatible arguments/returns.
Generating WAT from Wasm ^^^^^^^^^^^^^^^^^^^^^^^^
For those who want to use precompiled Wasm modules, it's enough to translate Wasm bytecode to wat representation. On Linux, it can be achieved by a wasm2wat tool, available in most distributions in the wabt package.
Here's how a wasm function can be declared:
.. code-block:: cql
CREATE FUNCTION ks.fib (input bigint) RETURNS NULL ON NULL INPUT RETURNS bigint LANGUAGE wasm AS '
(module
(func $fib (param $n i64) (result i64)
(if
(i64.lt_s (local.get $n) (i64.const 2))
(return (local.get $n))
)
(i64.add
(call $fib (i64.sub (local.get $n) (i64.const 1)))
(call $fib (i64.sub (local.get $n) (i64.const 2)))
)
)
(memory (;0;) 2)
(export "memory" (memory 0))
(export "fib" (func $fib))
(global (;0;) i32 (i32.const 1024))
(export "_scylla_abi" (global 0))
(data $.rodata (i32.const 1024) "\01")
)'
and it can be invoked just like a regular UDF:
.. code-block:: cql
scylla@cqlsh:ks> CREATE TABLE t(id int, n bigint, PRIMARY KEY(id,n));
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 0);
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 1);
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 2);
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 3);
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 4);
scylla@cqlsh:ks> INSERT INTO t(id, n) VALUES (0, 5);
scylla@cqlsh:ks> SELECT n, ks.fib(n) FROM t;
n | ks.fib(n)
----+-----------
0 | 0
1 | 1
2 | 1
3 | 2
4 | 3
5 | 5
(6 rows)
WebAssembly UDFs are still experimental due to insufficient testing. If backwards incompatible changes
to the ABI are implemented in the future, they should be submitted as new ABI-versions, and use the same
LANGUAGE wasm clause in the CQL statements.
.. _abi-versions:
Different programming languages may require different ABIs. To support that, the Wasm program is required to
export the symbol _scylla_abi, that is a WebAssembly global with a 32-bit value of the offset in memory,
where the version number can be read (that's the only method of exporting a constant in Rust).
Currently, the only available ABI versions are 1 and 2. Both of them use the same protocol for passing parameters and returning values, but they differ in approaches to memory management.
The memory management differs depending on the used ABI version:
version 1 There are no requirements of the usage of memory by the user. The host grows memory for each of the parameters and does not free the memory in any way.
version 2
The user program is required to export _scylla_malloc and _scylla_free methods, which
are then used by the host for allocating memory for parameters and freeing memory for the returned value.
The user is required to free the memory allocated for parameters using the _scylla_free method, and allocate the memory for result using the _scylla_malloc method (both can be achieved by using the provided helper libraries).
Alternatively, the user may return one of the arguments, shifting the responsibility of freeing it to the host.
The _scylla_malloc and _scylla_free methods may be simple wrappers of malloc and free methods implemented by default when compiling with WASI.
Due to the limitations imposed by WebAssembly specification, the following types can be natively supported with CQL:
======================= =====================
CQL type Wasm type
======================= =====================
bigint i64
boolean i32
double f64
float f32
int i32
smallint i32
tinyint i32
======================= =====================
The rest of CQL types (text, date, timestamp, etc.) are implemented by putting their serialized representation into Wasm module memory and passing for each parameter a 64-bit value, of which top 32 bits are its size and its bottom 32 bits are a pointer to its serialized representation, like below:
.. code-block:: c
int32_t size = foo.size();
int32_t ptr = (int32_t)malloc(size);
int64_t param = ((int64_t)size << 32) | ptr;
Native WebAssembly types can only be represented directly if the function does not operate on NULL values. Fortunately, user-defined functions
explicitly specify whether they accept NULL or not.
If the function is specified not to accept NULL, all parameters and return values are represented
as in the description above.
If the function is specified to accept NULL, parameters and return values of both natively and non-natively supported types are represented
using their serialized representation, also described above.
The important distinction is that size equal to -1 (minus one or 0xffffffff) indicates that the value is NULL and should not be parsed.
.. note:: CQL syntax extensions and new helper libraries may be deployed together with new ABI versions. The description below only refers to ABI versions 1 and 2.
Currently, returning NULL values is possible only for functions declared to be CALLED ON NULL INPUT.
This decision allows returning some values as native WebAssembly types without having to allocate memory for them and serialize them first.
Alternative ways of expressing whether a function can return null should be considered - perhaps as CQL syntax extension.