Re: [WIP] [B-Tree] Retail IndexTuple deletion

From: Dmitry Dolgov <9erthalion6(at)gmail(dot)com>
To: a(dot)lepikhov(at)postgrespro(dot)ru
Cc: PostgreSQL-development <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: [WIP] [B-Tree] Retail IndexTuple deletion
Date: 2018-11-29 13:27:50
Message-ID: CA+q6zcV9rkCdZbt4DDyUXCZupB_2vLH0NKtCF6iR=dganGP_vg@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox
Thread:
Lists: pgsql-hackers

> On Fri, Sep 21, 2018 at 5:52 AM Andrey Lepikhov <a(dot)lepikhov(at)postgrespro(dot)ru> wrote:
>
> The v6 version of quick vacuum, which utilizes the amtargetdelete()
> interface for retail indextuple deletion.
> Now it is more simple and laconic.
> It must be applied after Retail-IndexTuple-Deletion-Access-Method.patch.
>
> BENCHMARKS:
> -----------
>
> Initial script:
> pgbench -i -s $scale # initial DB generation
> "CREATE INDEX pgbench_accounts_ext ON public.pgbench_accounts USING
> btree (abalance);" # additional index
>
> Comparison with lazy vacuum:
>
> script:
> "DELETE FROM pgbench_accounts WHERE (random() < $factor);" # delete a
> part of tuples for cleaning strategies comparison
> "VACUUM pgbench_accounts;" # check time of vacuum process by bash 'date
> +%s%N | cut -b1-13' command
>
> Results:
> | $scale=10 | $scale=100 |
> $factor| QVAC | LVAC | QVAC | LVAC |
> 1E-6 | - | - | 284 | 979 |
> 1E-5 | 78 | 144 | 288 | 1423 |
> 1E-4 | 72 | 280 | 388 | 3304 |
> 1E-3 | 189 | 609 | 2294 | 6029 |
> 1E-2 | 443 | 783 | 54232| 67884|
> 1E-1 | 1593 | 1237 | 83092| 86104|
>
> where QVAC - forced use of quick vacuum; LVAC - use lazy vacuum for
> index cleanup. $factor corresponds a number of vacuumed tuples. For
> example, $scale=10, $factor=1E-1 -> 100000 tuples vacuumed. Time
> measured in ms.
>
> So, quick strategy can be used in a vacuum process effectively up to
> 1-2% of DEAD tuples in a relation.

Hi,

Unfortunately, this patch doesn't compile anymore:

index.c: In function ‘htid2IndexDatum’:
index.c:4172:2: error: too few arguments to function ‘MakeSingleTupleTableSlot’
TupleTableSlot *slot = MakeSingleTupleTableSlot(RelationGetDescr(hrel));
^

Also I'm a bit confused about the current status of collaboration between this
patch and [1], one is tightly depends on another or not? Does it makes sense
to have only one corresponding CF item then? For now I'll move this one to
the next CF.

[1]: https://www.postgresql.org/message-id/flat/CAH2-WzkVb0Kom%3DR%2B88fDFb%3DJSxZMFvbHVC6Mn9LJ2n%3DX%3DkS-Uw%40mail.gmail.com

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message Stephen Frost 2018-11-29 13:28:36 Re: [PATCH] Change "checkpoint starting" message to use "wal"
Previous Message Alvaro Herrera 2018-11-29 13:27:07 Re: New GUC to sample log queries