Re: Partitioning and unique key - Mailing list pgsql-general

From veem v
Subject Re: Partitioning and unique key
Date
Msg-id CAB+=1TV+Pvm8N-24GWhoG+baDEeiCytQ5JnVze31D60HUyP95g@mail.gmail.com
Whole thread Raw
In response to Re: Partitioning and unique key  (veem v <veema0000@gmail.com>)
List pgsql-general


On Sun, 1 Sept 2024 at 11:38, veem v <veema0000@gmail.com> wrote:

On Sun, 1 Sept 2024 at 10:03, veem v <veema0000@gmail.com> wrote:

On Sun, 1 Sept 2024 at 09:13, David G. Johnston <david.g.johnston@gmail.com> wrote:
On Saturday, August 31, 2024, veem v <veema0000@gmail.com> wrote:


iii)And then alter the datatype of the partition key transaction_date to DATE in one shot at the table level(which should be fast as its having more granularity as compare to existing timestamptype, so should be catalog or dictionary change only), and that will remain the part of composite PK (transaction_id,transaction_date).

While this might seem logical, in reality date and timestamptz are different fixed-width data types and thus any attempt to change from one to the other will involve a table rewrite.  Best you could do is leave the timestamptz in place and just truncate to day so the time is always midnight UTC.
 

 
Here  , if we keep the PK column as is i.e. the transaction_timestamp as timestamptz but truncate the time component , in that case again in future if someone tries to insert(using insert on conflict) data into the table with time component , it will get consumed and will not be restricted by the PK constraint. So I was trying to make the data type also as DATE for the transaction_timestap column.

As in this case anyway we have to create another column to populate the date+timestamp values as we cant throw those values away per business need, so we will be kind of rewriting the table.So is it okay if if we will

1) Detach all the partitions.
2)Do the alter using "only" key word in table level. (For adding new column transaction_timestamp_new  to hold date+timestamp value and also altering the existing transaction_timestamp column to DATE from type timestamptz).
3)Then do the data fix(delete the duplicates) and alter the column, one partition at a time for all of the partitions and once done , attach  those partitions one by one.
5)Rename the columns at table level.Hope this won't need any table rewrite.

Is there any downside if we go by the above approach?

Or do you mean to say there is no way we can modify the data type of a partition key even by detaching the partitions one by one? And thus we may have only way left is to create the table from scratch with partitions and populate the data to it? I was avoiding this because we have many indexes also in it , so creating from scratch means creating those indexes again. So I wanted to achieve it by detaching partitions, doing the required change and attaching it again.


I tried  by detaching the partitions, but still then it's not allowing me to alter the DAT TYPE of the partition key and throwing error as below. 

ERROR:  cannot alter column "<patition_key_column>" because it is part of the partition key of relation "<table_name>"

Now I am thinking if it's really going to get too complex if we try to stick with the partition detach and attach strategy. As a few teammates say , having a new column added with just a date type and then drop the existing FK and PK first and then detach all the partitions, and attach the partitions back using the new DATE column. and then recreate the PK again. Btw we have ~5 partition tables with parent child relationship on which this fix has to be applied. So I'm still wondering the best way possible for fixing this issue.

pgsql-general by date:

Previous
From: veem v
Date:
Subject: Re: Partitioning and unique key
Next
From: Adrian Klaver
Date:
Subject: Re: Upgrade Ubuntu 22 -> 24 may break PostgreSQL