Re: advice sought - general approaches to optimizing queries around "event streams"

We have a similar timeseries database approaching 500m records.

We partition the main tables (much like your events) into one year subsets, with a clustered index on timestamp for all but the live year.

https://blog.engineyard.com/2013/scaling-postgresql-performance-table-partitioning
http://www.postgresql.org/docs/9.3/static/ddl-partitioning.html
http://www.postgresql.org/docs/9.3/static/sql-cluster.html

As discussed here previously, you can also improve performance using hardware - SSD'd vs spindles. Also note that tablespaces, with indexes on your faster drives & data on slower ones can improve performance.

http://www.postgresql.org/docs/9.3/static/manage-ag-tablespaces.html

Also make sure your db server is optimised for the database size & hardware configuration - like perhaps alloe fewer concurrent users, but more resources per user, or see what pgtune recommends.

Should help your performance, in terms of underlying db efficiency & performance, rather than tweaking your actual queries.

Brent Wood

Programme leader: Environmental Information Delivery
NIWA
DDI:  +64 (4) 3860529

________________________________________
From: pgsql-general-owner@postgresql.org <pgsql-general-owner@postgresql.org> on behalf of Jonathan Vanasco <postgres@2xlp.com>
Sent: Saturday, September 27, 2014 9:02 AM
To: PostgreSQL general
Subject: [GENERAL] advice sought - general approaches to optimizing queries around "event streams"

I have a growing database with millions of rows that track resources against an event stream.

i have a few handfuls of queries that interact with this stream in a variety of ways, and I have managed to drop things down from 70s to 3.5s on full scans and offer .05s partial scans.

no matter how i restructure queries, I can't seem to get around a few bottlenecks and I wanted to know if there were any tips/tricks from the community on how to approach them.

a simple form of my database would be:

       --  1k of
       create table stream (
               id int not null primary key,
       )

       -- 1MM of
       create table resource (
               id int not null primary key,
               col_a bool,
               col_b bool,
               col_c text,
       );

       -- 10MM of
       create table streamevent (
               id int not null,
               event_timestamp timestamp not null,
               stream_id int not null references stream(id)
       );

       -- 10MM of
       create table resource_2_stream_event(
               resource_id int not null references resource(id),
               streamevent_id int not null references streamevent(id)
       )

Everything is running off of indexes; there are no seq scans.

I've managed to optimize my queries by avoiding joins against tables, and turning the stream interaction into a subquery or CTE.
better performance has come from limiting the number of "stream events"  ( which are only the timestamp and resource_id off a joined table )

The bottlenecks I've encountered have primarily been:

1.      When interacting with a stream, the ordering of event_timestamp and deduplicating of resources becomes an issue.
       I've figured out a novel way to work with the most recent events, but distant events are troublesome

       using no limit, the query takes 3500 ms
       using a limit of 10000, the query takes 320ms
       using a limit of 1000, the query takes 20ms

       there is a dedicated index of on event_timestamp (desc) , and it is being used
       according to the planner... finding all the records is fine; merging-into and sorting the aggregate to handle the deduplication of records in a stream seems to be the issue (either with DISTINCT or max+group_by)


2.      I can't figure out an effective way to search for a term against an entire stream (using a tsquery/gin based search)

       I thought about limiting the query by finding matching resources first, then locking it to an event stream, but:
               - scanning the entire table for a term takes about 10 seconds on an initial hit.  subsequent queries for the same terms end up using the cache, and complete within 20ms.

       I get better search performance by calculating the event stream, then searching it for matching documents, but I still have the performance issues related to limiting the window of events

i didn't include example queries, because I'm more concerned with the general approaches and ideas behind dealing with large data sets than i am with raw SQL right now.

i'm hoping someone can enlighten me into looking at new ways to solve these problems.   i think i've learned more about postgres/sql in the past 48hour than I have in the past 15 years, and I'm pretty sure that the improvements I need will come from new ways of querying data , rather than optimizing the current queries.




















--
Sent via pgsql-general mailing list (pgsql-general@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-general