Обсуждение: Performance decline maybe caused by multi-column index?

Hi,

While developing a batch processing platform using postgresql as the
underlying data store we are seeing a performance decline in our
application.

In this application a job is broken up into chunks where each chunk
contains a number of items (typically 10).

CREATE TABLE item (
     id                      SMALLINT NOT NULL,
     chunkId            INTEGER NOT NULL,
     jobId                 INTEGER NOT NULL,
     -- other attributes omitted for brewity
     PRIMARY KEY (jobId, chunkId, id)
);

So a job with 600.000 items results in 600.000 rows in the items table
with a fixed jobId, chunkId ranging from 0-59999 and for each chunkId an
id ranging from 0-9.

All ten inserts for a particular chunkId are handled in a single
transaction, and over time we are seeing an increase in transaction
execution time, <100ms for the first 100.000 items, >300ms when we reach
the 400.000 mark, and the trend seems to be forever increasing.

No decline is observed if we instead sequentially submit 6 jobs of
100.000 items each.

Therefore we are beginning to wonder if we are hitting some sort of
upper limit with regards to the multi column index? Perhaps something
causing it to sort on disk or something like that?

Any suggestions to the cause of this would be very much appreciated.

jobstore=> SELECT version();
                                            version
----------------------------------------------------------------------------------------------
  PostgreSQL 9.4.4 on x86_64-unknown-linux-gnu, compiled by gcc (Debian
4.7.2-5) 4.7.2, 64-bit

jobstore=> SELECT name, current_setting(name), SOURCE
jobstore->   FROM pg_settings
jobstore->   WHERE SOURCE NOT IN ('default', 'override');
             name            |            current_setting |        source
----------------------------+----------------------------------------+----------------------
  application_name           | psql | client
  client_encoding            | UTF8 | client
  DateStyle                  | ISO, YMD | configuration file
  default_text_search_config | pg_catalog.english | configuration file
  lc_messages                | en_DK.UTF-8 | configuration file
  lc_monetary                | en_DK.UTF-8 | configuration file
  lc_numeric                 | en_DK.UTF-8 | configuration file
  lc_time                    | en_DK.UTF-8 | configuration file
  listen_addresses           | * | configuration file
  log_line_prefix            | %t | configuration file
  log_timezone               | localtime | configuration file
  max_connections            | 100 | configuration file
  max_stack_depth            | 2MB | environment variable
  port                       | 5432 | configuration file
  shared_buffers             | 128MB | configuration file
  ssl                        | on | configuration file
  ssl_cert_file              | /etc/ssl/certs/ssl-cert-snakeoil.pem |
configuration file
  ssl_key_file               | /etc/ssl/private/ssl-cert-snakeoil.key |
configuration file
  TimeZone                   | localtime | configuration file

Kind regards,

Jan Bauer Nielsen
Software developer
DBC as
http://www.dbc.dk/english

On Fri, Mar 18, 2016 at 6:26 AM, Jan Bauer Nielsen <jbn@dbc.dk> wrote:
> Hi,
>
> While developing a batch processing platform using postgresql as the
> underlying data store we are seeing a performance decline in our
> application.
>
> In this application a job is broken up into chunks where each chunk contains
> a number of items (typically 10).
>
> CREATE TABLE item (
>     id                      SMALLINT NOT NULL,
>     chunkId            INTEGER NOT NULL,
>     jobId                 INTEGER NOT NULL,
>     -- other attributes omitted for brewity
>     PRIMARY KEY (jobId, chunkId, id)
> );
>
> So a job with 600.000 items results in 600.000 rows in the items table with
> a fixed jobId, chunkId ranging from 0-59999 and for each chunkId an id
> ranging from 0-9.

Is it 0-59999 in order, or in some arbitrary order?

>
> All ten inserts for a particular chunkId are handled in a single
> transaction, and over time we are seeing an increase in transaction
> execution time, <100ms for the first 100.000 items, >300ms when we reach the
> 400.000 mark, and the trend seems to be forever increasing.

Why such small transactions?  Why not do the entire 600.000 in on transaction?

Are you inserting them via COPY, or doing single-valued inserts in a
loop, or inserts with multiple value lists?


>
> No decline is observed if we instead sequentially submit 6 jobs of 100.000
> items each.
>
> Therefore we are beginning to wonder if we are hitting some sort of upper
> limit with regards to the multi column index? Perhaps something causing it
> to sort on disk or something like that?


My gut feeling is that is more about memory management in your client,
rather than something going on in the database.  What does `top`, or
`perf top`, show you about what is going on?

Can you produce a simple perl or python script that reproduces the problem?

Cheers,

Jeff