Обсуждение: Schema for caching message-count in folders using triggers
Hi all.
I'm facing a problem with my current schema for email where folders start containing several 100K of messages and count(*) in them taks noticeable time. This schema is accessible from IMAP and a web-app so lots of queries of the type "list folders with message count" are performed.
So, I'm toying with this idea of caching the message-count in the folder-table itself.
I currently have this:
CREATE or replace FUNCTION count_increment_tf() RETURNS TRIGGER AS $_$
BEGIN
UPDATE folder SET message_count = message_count + 1 WHERE id = NEW.folder_id;
RETURN NEW;
END $_$ LANGUAGE 'plpgsql';
CREATE or replace FUNCTION count_decrement_tf() RETURNS TRIGGER AS $_$
BEGIN
UPDATE folder SET message_count = message_count - 1 WHERE id = OLD.folder_id;
RETURN OLD;
END $_$ LANGUAGE 'plpgsql';
CREATE or replace FUNCTION count_update_tf() RETURNS TRIGGER AS $_$
BEGIN
UPDATE folder SET message_count = message_count - 1 WHERE id = OLD.folder_id;
UPDATE folder SET message_count = message_count + 1 WHERE id = NEW.folder_id;
RETURN NEW;
END $_$ LANGUAGE 'plpgsql';
CREATE TRIGGER increment_folder_msg_t AFTER INSERT ON message FOR EACH ROW EXECUTE PROCEDURE count_increment_tf();
CREATE TRIGGER decrement_folder_msg_t AFTER DELETE ON message FOR EACH ROW EXECUTE PROCEDURE count_decrement_tf();
CREATE TRIGGER update_folder_msg_t AFTER UPDATE ON message FOR EACH ROW EXECUTE PROCEDURE count_update_tf();
The problem with this is locking (waiting for another TX to commit when updating the same folder) and deadlock issues when trying to simultaneously insert/delete/update messages in a folder.
Does anyone have any better ideas for safely caching the message-count in each folder without locking and deadlock issues?
Thanks.
-- 
Andreas Joseph Krogh
CTO / Partner - Visena AS
Mobile: +47 909 56 963
Вложения
På torsdag 05. mars 2015 kl. 20:59:28, skrev Adrian Klaver <adrian.klaver@aklaver.com>:
On 03/05/2015 11:45 AM, Andreas Joseph Krogh wrote:
> Hi all.
> I'm facing a problem with my current schema for email where folders
> start containing several 100K of messages and count(*) in them taks
> noticeable time. This schema is accessible from IMAP and a web-app so
> lots of queries of the type "list folders with message count" are performed.
> So, I'm toying with this idea of caching the message-count in the
> folder-table itself.
> I currently have this:
>
> CREATE or replace FUNCTIONcount_increment_tf()RETURNS TRIGGER AS$_$
> BEGIN
> UPDATE folder SETmessage_count =message_count + 1 WHEREid =NEW.folder_id;
> RETURNNEW;
> END $_$LANGUAGE'plpgsql';
>
> CREATE or replace FUNCTIONcount_decrement_tf()RETURNS TRIGGER AS$_$
> BEGIN
> UPDATE folder SETmessage_count =message_count - 1 WHEREid =OLD.folder_id;
> RETURNOLD;
> END $_$LANGUAGE'plpgsql';
>
> CREATE or replace FUNCTIONcount_update_tf()RETURNS TRIGGER AS$_$
> BEGIN
> UPDATE folder SETmessage_count =message_count - 1 WHEREid =OLD.folder_id;
> UPDATE folder SETmessage_count =message_count + 1 WHEREid =NEW.folder_id;
> RETURNNEW;
> END $_$LANGUAGE'plpgsql';
>
> CREATE TRIGGERincrement_folder_msg_tAFTER INSERT ON message FOR EACH ROW EXECUTE PROCEDUREcount_increment_tf();
> CREATE TRIGGERdecrement_folder_msg_tAFTER DELETE ON message FOR EACH ROW EXECUTE PROCEDUREcount_decrement_tf();
> CREATE TRIGGERupdate_folder_msg_tAFTER UPDATE ON message FOR EACH ROW EXECUTE PROCEDUREcount_update_tf();
>
> The problem with this is locking (waiting for another TX to commit when
> updating the same folder) and deadlock issues when trying to
> simultaneously insert/delete/update messages in a folder.
> Does anyone have any better ideas for safely caching the message-count
> in each folder without locking and deadlock issues?
How accurate does this have to be?
Not exactly following what is folder?
Is it a table that contains the messages?
A top of the head idea would be to use sequences. Create a sequence for
each folder starting at current count and then use nextval, setval to
change the value:
http://www.postgresql.org/docs/9.4/interactive/functions-sequence.html
It is not transactional, so it would probably not be spot on, which is
why I asked about accuracy earlier.
Yes, 'folder' is a table which contains 'message':
create table folder( id serial PRIMARY KEY, name varchar not null unique, message_count integer not null default 0
);
create table message( id serial PRIMARY KEY, folder_id INTEGER NOT NULL REFERENCES folder(id), message varchar not null
);
The count has to be exact, no estimate from EXPLAIN or such...
-- 
Andreas Joseph Krogh
CTO / Partner - Visena AS
Mobile: +47 909 56 963
Вложения
På torsdag 05. mars 2015 kl. 22:16:01, skrev Alvaro Herrera <alvherre@2ndquadrant.com>:
Andreas Joseph Krogh wrote:
> Hi all. I'm facing a problem with my current schema for email where folders
> start containing several 100K of messages and count(*) in them taks noticeable
> time. This schema is accessible from IMAP and a web-app so lots of queries of
> the type "list folders with message count" are performed. So, I'm toying with
> this idea of caching the message-count in the folder-table itself.
You can do this better by keeping a table with per-folder counts and
deltas. There is one main row which keeps the total value at some point
in time. Each time you insert a message, add a "delta" entry with value
1; each time you remove, add a delta with value -1. You can do this
with a trigger on insert/update/delete. This way, there is no
contention because there are no updates.
To figure out the total value, just add all the values (the main plus
all deltas for that folder).
>From time to time you have a process that summarizes all these entries
into one total value again. Something like
WITH deleted AS (DELETE
FROM counts
WHERE type = 'delta' RETURNING value),
total AS (SELECT coalesce(sum(value), 0) as sum
FROM deleted)
UPDATE counts
SET value = counts.value + total.sum
FROM total WHERE type = 'total'
RETURNING counts.value
Like it, thanks!
-- 
Andreas Joseph Krogh
CTO / Partner - Visena AS
Mobile: +47 909 56 963