Re: a JOIN to a VIEW seems slow

["Frank Millman" <frank@xxxxxxxxxxxx>]

 

2017-09-14 15:09 GMT+02:00 Pavel Stehule <pavel.stehule@xxxxxxxxx>:

 

 

2017-09-14 14:59 GMT+02:00 Frank Millman <frank@xxxxxxxxxxxx>:

Pavel Stehule wrote:

 

2017-09-14 10:14 GMT+02:00 Frank Millman <frank@xxxxxxxxxxxx>:

Hi all

 

This is a follow-up to a recent question I posted regarding a slow query. I thought that the slowness was caused by the number of JOINs in the query, but with your assistance I have found the true reason. I said in the previous thread that the question had become academic, but now that I understand things better, it is no longer academic as it casts doubt on my whole approach.

 

I have split my AR transaction table into three physical tables – ar_tran_inv, ar_tran_crn, ar_tran_rec. I will probably add others at some point, such as ar_tran_jnl.

 

I then create a VIEW to view all transactions combined. The view is created like this -

 

CREATE VIEW ar_trans AS

  SELECT ‘ar_inv’ AS tran_type, row_id AS tran_row_id, tran_number ... FROM ar_tran_inv WHERE posted = ‘1’

  UNION ALL

  SELECT ‘ar_crn’ AS tran_type, row_id AS tran_row_id, tran_number ... FROM ar_tran_crn WHERE posted = ‘1’

  UNION ALL

  SELECT ‘ar_rec’ AS tran_type, row_id AS tran_row_id, tran_number ... FROM ar_tran_rec WHERE posted = ‘1’

 

I have another table called ‘ar_trans_due’, to keep track of outstanding transactions. All of the three transaction types generate entries into this table. To identify the source of the transaction, I have created columns in ar_trans_due called ‘tran_type’ and ‘tran_row_id’. After inserting a row into ‘ar_tran_inv’, I invoke this -

 

  INSERT INTO ar_trans_due (tran_type, tran_row_id, ...) VALUES (‘ar_inv’, ar_tran_inv.row_id, ...), and similar for the other transaction types. It is handled by a Python program, and it all happens within a transaction.

 

When I view a row in ar_trans_due, I want to retrieve data from the source transaction, so I have this -

 

  SELECT * FROM ar_trans_due a

  LEFT JOIN ar_trans b ON b.tran_type = a.tran_type AND b.tran_row_id = a.tran_row_id

 

I understand that PostgreSQL must somehow follow a path from the view ‘ar_trans’ to the physical table ‘ar_tran_inv’, but I assumed it would execute the equivalent of SELECT * FROM ar_tran_inv WHERE row_id = a.tran_row_id AND posted = ‘1’.

 

If this was the case, it would be an indexed read, and very fast. Instead, according to EXPLAIN, it performs a sequential scan of the ‘ar_tran_inv’ table.

 

It also scans ‘ar_tran_crn’ and ‘ar_tran_rec’, but EXPLAIN shows that it uses a Bitmap Heap Scan on those. I assume that is because the tables are currently empty.

 

Is this analysis correct?

 

please, send EXPLAIN ANALYZE result :)

 

 

I tried to reduce this to its simplest form.

 

Here is a SQL statement -

 

    SELECT *

    FROM ccc.ar_trans_due a

    LEFT JOIN ccc.ar_trans b ON

        b.tran_type = a.tran_type AND b.tran_row_id = a.tran_row_id

    WHERE a.row_id = 1

 

ar_trans_due is a physical table, ar_trans is a view.

 

It takes about 28ms. Here is the explain - https://explain.depesz.com/s/8YY

 

 

> The PostgreSQL cannot to push join - in slow case, the UNIONS should be done first - and it requires full scan ar_tran_inv - used filter (posted AND (deleted_id = 0) is not too effective - maybe some composite or partial index helps.

>

> The fast query doesn't contains unions - so there are bigger space for optimizer - ar_tran_inv is filtered effective - by primary key.

>

> So main problem is impossible to push information a.row_id = 1 to deep to query.

>

 

Sorry for banging on about this, but someone might be interested in the following timings.

 

The only solution I could find was to ‘denormalise’ (if that is a word) and create additional columns on ar_trans_due for cust_row_id and tran_date, to avoid using any joins.

 

Once I had done that, I could run my query two ways –

  1. using the newly created columns

  2. as before, using a join to the view, which in turn retrieved data from the underlying tables.

 

This was a more complex query than the example above – details available on request.

 

Here are the timings for running the query on identical data sets using Postgresql, Sql Server, and Sqlite3 -

 

PostgreSQL -

    Method 1 - 0.28 sec

    Method 2 – 1607 sec, or 26 minutes

 

Sql Server -

    Method 1 – 0.33 sec

    Method 2 – 1.8 sec

 

Sqlite3 -

    Method 1 – 0.15 sec

    Method 2 – 1.0 sec

 

It seems that Sql Server and Sqlite3 are able to analyse the ‘join’, and execute an indexed read against the underlying physical tables.

 

Frank