[Tarantool-patches] [PATCH v2 09/15] sql: rework AVG()

[Mergen Imeev via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Thank you for the review! My answers, diff and new patch below.

On 15.09.2021 00:24, Vladislav Shpilevoy wrote:
> Thanks for the patch!
>
>> diff --git a/src/box/sql/func.c b/src/box/sql/func.c
>> index 12a6a5a2c..9e0c09206 100644
>> --- a/src/box/sql/func.c
>> +++ b/src/box/sql/func.c
>> @@ -102,6 +102,44 @@ fin_total(struct sql_context *ctx)
>>  		mem_copy_as_ephemeral(ctx->pOut, ctx->pMem);
>>  }
>>  
>> +/** Implementation of the AVG() function. */
>> +static void
>> +step_avg(struct sql_context *ctx, int argc, struct Mem **argv)
>> +{
>> +	assert(argc == 1);
>> +	(void)argc;
>> +	assert(ctx->pMem->type == MEM_TYPE_NULL || mem_is_bin(ctx->pMem));
>> +	if (argv[0]->type == MEM_TYPE_NULL)
>> +		return;
>> +	if (ctx->pMem->type == MEM_TYPE_NULL) {
>> +		uint32_t size = 2 * sizeof(struct Mem);
>> +		struct Mem *mems = sqlDbMallocRawNN(sql_get(), size);
>
> Previously only size of mem + uint32 was allocated, now it is 2 size
> of mem. Lets stick to the more compact version. For the division in the
> end you can create a mem on the stack, for instance.
Thanks, fixed.

Diff:

diff --git a/src/box/sql/func.c b/src/box/sql/func.c
index 9e0c09206..b5f154fb1 100644
--- a/src/box/sql/func.c
+++ b/src/box/sql/func.c
@@ -111,20 +111,26 @@ step_avg(struct sql_context *ctx, int argc, struct Mem **argv)
 	assert(ctx->pMem->type == MEM_TYPE_NULL || mem_is_bin(ctx->pMem));
 	if (argv[0]->type == MEM_TYPE_NULL)
 		return;
+	struct Mem *mem;
+	uint32_t *count;
 	if (ctx->pMem->type == MEM_TYPE_NULL) {
-		uint32_t size = 2 * sizeof(struct Mem);
-		struct Mem *mems = sqlDbMallocRawNN(sql_get(), size);
-		mem_create(&mems[0]);
-		mem_create(&mems[1]);
-		mem_copy_as_ephemeral(&mems[0], argv[0]);
-		mem_set_uint(&mems[1], 1);
-		mem_set_bin_allocated(ctx->pMem, (char *)mems, size);
+		uint32_t size = sizeof(struct Mem) + sizeof(uint32_t);
+		mem = sqlDbMallocRawNN(sql_get(), size);
+		if (mem == NULL) {
+			ctx->is_aborted = true;
+			return;
+		}
+		count = (uint32_t *)(mem + 1);
+		mem_create(mem);
+		*count = 1;
+		mem_copy_as_ephemeral(mem, argv[0]);
+		mem_set_bin_allocated(ctx->pMem, (char *)mem, size);
 		return;
 	}
-	struct Mem *mems = (struct Mem *)ctx->pMem->z;
-	assert(mems[1].type = MEM_TYPE_UINT);
-	++mems[1].u.u;
-	if (mem_add(&mems[0], argv[0], &mems[0]) != 0)
+	mem = (struct Mem *)ctx->pMem->z;
+	count = (uint32_t *)(mem + 1);
+	++*count;
+	if (mem_add(mem, argv[0], mem) != 0)
 		ctx->is_aborted = true;
 }
 
@@ -135,8 +141,12 @@ fin_avg(struct sql_context *ctx)
 	assert(ctx->pMem->type == MEM_TYPE_NULL || mem_is_bin(ctx->pMem));
 	if (ctx->pMem->type == MEM_TYPE_NULL)
 		return mem_set_null(ctx->pOut);
-	struct Mem *mems = (struct Mem *)ctx->pMem->z;
-	if (mem_div(&mems[0], &mems[1], ctx->pOut) != 0)
+	struct Mem *mem = (struct Mem *)ctx->pMem->z;
+	uint32_t *count = (uint32_t *)(mem + 1);
+	struct Mem mem_count;
+	mem_create(&mem_count);
+	mem_set_uint(&mem_count, *count);
+	if (mem_div(mem, &mem_count, ctx->pOut) != 0)
 		ctx->is_aborted = true;
 }
 

New patch:

commit 48ee23c47c696c5e002347841b652677ca877507
Author: Mergen Imeev <imeevma@gmail.com>
Date:   Thu Sep 9 18:19:53 2021 +0300

    sql: rework AVG()
    
    This patch makes AVG() accept DOUBLE values by default. Also, after this
    patch AVG() will be able to work with DECIMAL values.
    
    Part of #4145
    Part of #6355

diff --git a/src/box/sql/func.c b/src/box/sql/func.c
index 12a6a5a2c..b5f154fb1 100644
--- a/src/box/sql/func.c
+++ b/src/box/sql/func.c
@@ -102,6 +102,54 @@ fin_total(struct sql_context *ctx)
 		mem_copy_as_ephemeral(ctx->pOut, ctx->pMem);
 }
 
+/** Implementation of the AVG() function. */
+static void
+step_avg(struct sql_context *ctx, int argc, struct Mem **argv)
+{
+	assert(argc == 1);
+	(void)argc;
+	assert(ctx->pMem->type == MEM_TYPE_NULL || mem_is_bin(ctx->pMem));
+	if (argv[0]->type == MEM_TYPE_NULL)
+		return;
+	struct Mem *mem;
+	uint32_t *count;
+	if (ctx->pMem->type == MEM_TYPE_NULL) {
+		uint32_t size = sizeof(struct Mem) + sizeof(uint32_t);
+		mem = sqlDbMallocRawNN(sql_get(), size);
+		if (mem == NULL) {
+			ctx->is_aborted = true;
+			return;
+		}
+		count = (uint32_t *)(mem + 1);
+		mem_create(mem);
+		*count = 1;
+		mem_copy_as_ephemeral(mem, argv[0]);
+		mem_set_bin_allocated(ctx->pMem, (char *)mem, size);
+		return;
+	}
+	mem = (struct Mem *)ctx->pMem->z;
+	count = (uint32_t *)(mem + 1);
+	++*count;
+	if (mem_add(mem, argv[0], mem) != 0)
+		ctx->is_aborted = true;
+}
+
+/** Finalizer for the AVG() function. */
+static void
+fin_avg(struct sql_context *ctx)
+{
+	assert(ctx->pMem->type == MEM_TYPE_NULL || mem_is_bin(ctx->pMem));
+	if (ctx->pMem->type == MEM_TYPE_NULL)
+		return mem_set_null(ctx->pOut);
+	struct Mem *mem = (struct Mem *)ctx->pMem->z;
+	uint32_t *count = (uint32_t *)(mem + 1);
+	struct Mem mem_count;
+	mem_create(&mem_count);
+	mem_set_uint(&mem_count, *count);
+	if (mem_div(mem, &mem_count, ctx->pOut) != 0)
+		ctx->is_aborted = true;
+}
+
 static const unsigned char *
 mem_as_ustr(struct Mem *mem)
 {
@@ -1656,69 +1704,6 @@ soundexFunc(sql_context * context, int argc, sql_value ** argv)
 	}
 }
 
-/*
- * An instance of the following structure holds the context of a
- * sum() or avg() aggregate computation.
- */
-typedef struct SumCtx SumCtx;
-struct SumCtx {
-	struct Mem mem;
-	uint32_t count;
-};
-
-/*
- * Routines used to compute the sum, average, and total.
- *
- * The SUM() function follows the (broken) SQL standard which means
- * that it returns NULL if it sums over no inputs.  TOTAL returns
- * 0.0 in that case.  In addition, TOTAL always returns a float where
- * SUM might return an integer if it never encounters a floating point
- * value.  TOTAL never fails, but SUM might through an exception if
- * it overflows an integer.
- */
-static void
-sum_step(struct sql_context *context, int argc, sql_value **argv)
-{
-	assert(argc == 1);
-	UNUSED_PARAMETER(argc);
-	struct SumCtx *p = sql_aggregate_context(context, sizeof(*p));
-	if (p == NULL) {
-		context->is_aborted = true;
-		return;
-	}
-	if (p->count == 0) {
-		mem_create(&p->mem);
-		assert(context->func->def->returns == FIELD_TYPE_INTEGER ||
-		       context->func->def->returns == FIELD_TYPE_DOUBLE);
-		if (context->func->def->returns == FIELD_TYPE_INTEGER)
-			mem_set_uint(&p->mem, 0);
-		else
-			mem_set_double(&p->mem, 0.0);
-	}
-	if (argv[0]->type == MEM_TYPE_NULL)
-		return;
-	++p->count;
-	assert(mem_is_num(argv[0]));
-	if (mem_add(&p->mem, argv[0], &p->mem) != 0)
-		context->is_aborted = true;
-}
-
-static void
-avgFinalize(sql_context * context)
-{
-	SumCtx *p;
-	p = sql_aggregate_context(context, 0);
-	if (p == NULL || p->count == 0) {
-		mem_set_null(context->pOut);
-		return;
-	}
-	struct Mem mem;
-	mem_create(&mem);
-	mem_set_uint(&mem, p->count);
-	if (mem_div(&p->mem, &mem, context->pOut) != 0)
-		context->is_aborted = true;
-}
-
 /*
  * The following structure keeps track of state information for the
  * count() aggregate function.
@@ -2015,8 +2000,9 @@ struct sql_func_definition {
 static struct sql_func_definition definitions[] = {
 	{"ABS", 1, {FIELD_TYPE_INTEGER}, FIELD_TYPE_INTEGER, absFunc, NULL},
 	{"ABS", 1, {FIELD_TYPE_DOUBLE}, FIELD_TYPE_DOUBLE, absFunc, NULL},
-	{"AVG", 1, {FIELD_TYPE_INTEGER}, FIELD_TYPE_INTEGER, sum_step, avgFinalize},
-	{"AVG", 1, {FIELD_TYPE_DOUBLE}, FIELD_TYPE_DOUBLE, sum_step, avgFinalize},
+	{"AVG", 1, {FIELD_TYPE_DOUBLE}, FIELD_TYPE_DOUBLE, step_avg, fin_avg},
+	{"AVG", 1, {FIELD_TYPE_INTEGER}, FIELD_TYPE_INTEGER, step_avg, fin_avg},
+	{"AVG", 1, {FIELD_TYPE_DECIMAL}, FIELD_TYPE_DECIMAL, step_avg, fin_avg},
 	{"CHAR", -1, {FIELD_TYPE_INTEGER}, FIELD_TYPE_STRING, charFunc, NULL},
 	{"CHAR_LENGTH", 1, {FIELD_TYPE_STRING}, FIELD_TYPE_INTEGER, lengthFunc,
 	 NULL},
diff --git a/test/sql-tap/built-in-functions.test.lua b/test/sql-tap/built-in-functions.test.lua
index 507d06549..08a63b86d 100755
--- a/test/sql-tap/built-in-functions.test.lua
+++ b/test/sql-tap/built-in-functions.test.lua
@@ -1,6 +1,6 @@
 #!/usr/bin/env tarantool
 local test = require("sqltester")
-test:plan(58)
+test:plan(61)
 
 local dec = require('decimal')
 
@@ -477,7 +477,7 @@ test:do_test(
         local res = {pcall(box.execute, [[SELECT AVG(?);]], {'1'})}
         return {tostring(res[3])}
     end, {
-        "Type mismatch: can not convert string('1') to integer"
+        "Type mismatch: can not convert string('1') to double"
     })
 
 test:do_catchsql_test(
@@ -605,4 +605,33 @@ test:do_execsql_test(
     }
 )
 
+-- Make sure AVG() accepts and returns DOUBLE by default.
+test:do_test(
+    "builtins-4.1.1",
+    function()
+        return box.execute([[SELECT AVG(?);]], {1}).metadata
+    end, {
+        {name = "COLUMN_1", type = "double"},
+    })
+
+test:do_test(
+    "builtins-4.1.2",
+    function()
+        local res = {pcall(box.execute, [[SELECT AVG(?);]], {-1ULL})}
+        return {tostring(res[3])}
+    end, {
+        "Type mismatch: can not convert integer(18446744073709551615) to double"
+    })
+
+-- Make sure AVG() works with DECIMAL properly.
+test:do_execsql_test(
+    "builtins-4.1.3",
+    [[
+        SELECT AVG(cast(column_2 as DECIMAL)) from (values(1), (123.432));
+    ]],
+    {
+        dec.new(62.216)
+    }
+)
+
 test:finish_test()