Обсуждение: Help with large delete

Currently I have one table that mimics a file system.  Each entry has a parent_id and a base name where parent_id is an
idin the table that must exist in the table or be null with cascade on delete. 

I’ve started a delete of a root entry with about 300,000 descendants.  The table currently has about 22M entries and
I’madding about 1600 entries per minute still.  Eventually there will not be massive amounts of entries being added and
thetable will be mostly static. 

I started the delete before from a terminal that got detached.  So I killed that process and started it up again from a
terminalless likely to get detached.˘ 

My question is basically how can I make life easier for Postgres?  I believe (hope) the deletes will be few and far
betweenbut they will happen from time to time.  In this case, Dropbox — its a long story that isn’t really pertinent.
Thepoint is that @#$% happens. 

“What can I do” includes starting completely over if necessary.  I’ve only got about a week invested in this and its
justmachine time at zero cost.  I could stop the other processes that are adding entries and let the delete finish if
thatwould help.  etc. 

Thank you for your time,
Perry

Currently I have one table that mimics a file system.  Each entry has a parent_id and a base name where parent_id is an id in the table that must exist in the table or be null with cascade on delete.

I’ve started a delete of a root entry with about 300,000 descendants.  The table currently has about 22M entries and I’m adding about 1600 entries per minute still.  Eventually there will not be massive amounts of entries being added and the table will be mostly static.

I started the delete before from a terminal that got detached.  So I killed that process and started it up again from a terminal less likely to get detached.˘

My question is basically how can I make life easier for Postgres?  I believe (hope) the deletes will be few and far between but they will happen from time to time.  In this case, Dropbox — its a long story that isn’t really pertinent.  The point is that @#$% happens.

“What can I do” includes starting completely over if necessary.  I’ve only got about a week invested in this and its just machine time at zero cost.  I could stop the other processes that are adding entries and let the delete finish if that would help.  etc.

Thank you for your time,
Perry

On 2022-04-16 08:25:56 -0500, Perry Smith wrote:
> Currently I have one table that mimics a file system.  Each entry has
> a parent_id and a base name where parent_id is an id in the table that
> must exist in the table or be null with cascade on delete.
>
> I’ve started a delete of a root entry with about 300,000 descendants.
> The table currently has about 22M entries and I’m adding about 1600
> entries per minute still.  Eventually there will not be massive
> amounts of entries being added and the table will be mostly static.
>
> I started the delete before from a terminal that got detached.  So I
> killed that process and started it up again from a terminal less
> likely to get detached.˘
>
> My question is basically how can I make life easier for Postgres?

Deleting 300k rows doesn't sound that bad. Neither does recursively
finding those 300k rows, although if you have a very biased distribution
(many nodes with only a few children, but some with hundreds of
thousands or even millions of children), PostgreSQL may not find a good
plan.

So as almost always when performance is an issue:

* What exactly are you doing?
* What is the execution plan?
* How long does it take?

        hp

--
   _  | Peter J. Holzer    | Story must make more sense than reality.
|_|_) |                    |
| |   | hjp@hjp.at         |    -- Charles Stross, "Creative writing
__/   | http://www.hjp.at/ |       challenge!"

Perry Smith <pedz@easesoftware.com> writes:
> Currently I have one table that mimics a file system.  Each entry has a parent_id and a base name where parent_id is
anid in the table that must exist in the table or be null with cascade on delete. 
> I’ve started a delete of a root entry with about 300,000 descendants.  The table currently has about 22M entries and
I’madding about 1600 entries per minute still.  Eventually there will not be massive amounts of entries being added and
thetable will be mostly static. 

The most obvious question is do you have an index on the referencing
column.  PG doesn't require one to exist to create an FK; but if you
don't, deletes of referenced rows had better be uninteresting to you
performance-wise, because each one will cause a seqscan.

            regards, tom lane

Perry Smith <pedz@easesoftware.com> writes:
> On Apr 16, 2022, at 10:33, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> The most obvious question is do you have an index on the referencing
>> column.  PG doesn't require one to exist to create an FK; but if you
>> don't, deletes of referenced rows had better be uninteresting to you
>> performance-wise, because each one will cause a seqscan.

> For Tom’s question, here is the description of the table:

> psql -c '\d dateien' find_dups
>                                           Table "public.dateien"
>    Column   |              Type              | Collation | Nullable |               Default
> ------------+--------------------------------+-----------+----------+-------------------------------------
>  id         | bigint                         |           | not null | nextval('dateien_id_seq'::regclass)
>  basename   | character varying              |           | not null |
>  parent_id  | bigint                         |           |          |
>  dev        | bigint                         |           | not null |
>  ftype      | character varying              |           | not null |
>  uid        | bigint                         |           | not null |
>  gid        | bigint                         |           | not null |
>  ino        | bigint                         |           | not null |
>  mode       | bigint                         |           | not null |
>  mtime      | timestamp without time zone    |           | not null |
>  nlink      | bigint                         |           | not null |
>  size       | bigint                         |           | not null |
>  sha1       | character varying              |           |          |
>  created_at | timestamp(6) without time zone |           | not null |
>  updated_at | timestamp(6) without time zone |           | not null |
> Indexes:
>     "dateien_pkey" PRIMARY KEY, btree (id)
>     "unique_dev_ino_for_dirs" UNIQUE, btree (dev, ino) WHERE ftype::text = 'directory'::text
>     "unique_parent_basename" UNIQUE, btree (COALESCE(parent_id, '-1'::integer::bigint), basename)
> Foreign-key constraints:
>     "fk_rails_c01ebbd0bf" FOREIGN KEY (parent_id) REFERENCES dateien(id) ON DELETE CASCADE
> Referenced by:
>     TABLE "dateien" CONSTRAINT "fk_rails_c01ebbd0bf" FOREIGN KEY (parent_id) REFERENCES dateien(id) ON DELETE CASCADE

Yeah.  So if you want to make deletes on this table not be unpleasantly
slow, you need an index on the parent_id column, and you don't have one.

(unique_parent_basename doesn't help, because with that definition it's
useless for looking up rows by parent_id.)

            regards, tom lane



On Apr 16, 2022, at 12:57, Jan Wieck <jan@wi3ck.info> wrote:

Make your connection immune to disconnects by using something like the screen utility.

Exactly… I’m using emacs in a server (daemon) mode so it stays alive.  Then I do “shell” within it.