Обсуждение: Is a clearer memory lifespan for outerTuple and innerTuple useful?

Hi,

When I am working on "shared detoast value"[0], where I want to avoid
detoast the same datum over and over again, I have to decide which
memory context should be used to hold the detoast value. later I 
found I have to use different MemoryContexts for the OuterTuple and
innerTuple since OuterTuple usually have a longer lifespan.

I found the following code in nodeMergeJoin.c which has pretty similar
situation, just that it uses ExprContext rather than MemoryContext.

MergeJoinState *
ExecInitMergeJoin(MergeJoin *node, EState *estate, int eflags)

    /*
     * we need two additional econtexts in which we can compute the join
     * expressions from the left and right input tuples.  The node's regular
     * econtext won't do because it gets reset too often.
     */
    mergestate->mj_OuterEContext = CreateExprContext(estate);
    mergestate->mj_InnerEContext = CreateExprContext(estate);

IIUC, we needs a MemoryContext rather than ExprContext in fact. In the
attachment, I just use two MemoryContext instead of the two ExprContexts
which should be less memory and more precise semantics, and works
fine. shall we go in this direction?  I attached the 2 MemoryContext in
JoinState rather than MergeJoinState, which is for the "shared detoast
value"[0] more or less.  

[0] https://www.postgresql.org/message-id/87ttoyihgm.fsf@163.com


-- 
Best Regards
Andy Fan

Andy Fan <zhihuifan1213@163.com> writes:

> ...,  I attached the 2 MemoryContext in
> JoinState rather than MergeJoinState, which is for the "shared detoast
> value"[0] more or less.  
>

After thinking more, if it is designed for "shared detoast value" patch
(happens on ExecInterpExpr stage), the inner_tuple_memory and
outer_tuple_memory should be attached to ExprContext rather than
JoinState since it is more natual to access ExprConext (compared with
JoinState) in ExecInterpExpr. I didn't attach a new version for this,
any feedback will be appreciated. 

-- 
Best Regards
Andy Fan

Andy Fan <zhihuifan1213@163.com> writes:

> Andy Fan <zhihuifan1213@163.com> writes:
>
>> ...,  I attached the 2 MemoryContext in
>> JoinState rather than MergeJoinState, which is for the "shared detoast
>> value"[0] more or less.  
>>

In order to delimit the scope of this discussion, I attached the 2
MemoryContext to MergeJoinState. Since the code was writen by Tom at
2005, so add Tom to the cc-list. 


-- 
Best Regards
Andy Fan

Nikita Malakhov <hukutoc@gmail.com> writes:

> Hi!
>
> Maybe, the alternative way is using a separate kind of context, say name it
> 'ToastContext' for all custom data related to Toasted values? What do
> you think?

That should be a candidate. The latest research makes me think the
'detoast_values' should have the same life cycles as tts_values, so the
memory should be managed by TupleTuleSlot (rather than ExprContext) and
be handled in ExecCopySlot / ExecClearSlot stuff.

In TupleTableSlot we already have a tts_mctx MemoryContext, reusing it
needs using 'pfree' to free the detoast values and but a dedicated
memory context pays more costs on the setup, but a more efficient
MemoryContextReset.

>
> On Sun, Dec 17, 2023 at 4:52 PM Andy Fan <zhihuifan1213@163.com> wrote:
>
>  Andy Fan <zhihuifan1213@163.com> writes:
>
>  > Andy Fan <zhihuifan1213@163.com> writes:
>  >
>  >> ...,  I attached the 2 MemoryContext in
>  >> JoinState rather than MergeJoinState, which is for the "shared detoast
>  >> value"[0] more or less.
>  >>
>
>  In order to delimit the scope of this discussion, I attached the 2
>  MemoryContext to MergeJoinState. Since the code was writen by Tom at
>  2005, so add Tom to the cc-list.
>

However this patch can be discussed seperately.

--
Best Regards
Andy Fan