Prepare for the PgBouncer and IPv4 deprecations on 26th January 2024

Learn more

What is new in PostgREST v11.1?

2023-07-12

6 minute read

PostgREST 11.1 is now available on the Supabase platform. Besides the pre-release features, we’ve added configuration and querying improvements. Here is what's new:

Impersonated Role Settings

Every role that passes PostgREST JWT Authentication is an impersonated role. On the Supabase platform, these are the anon, authenticated and service_role roles.

These roles can now have settings applied with a regular ALTER ROLE .. SET. This is useful, for example, to prevent web users from running expensive queries.

Let’s try it by setting a statement timeout and cost limit.

Statement timeout

statement_timeout aborts any statement that takes more than the specified amount of time. Let’s set it for the anon, authenticated and service_role roles:


_17
-- anonymous users can run queries that take 100 milliseconds max
_17
alter
_17
role anon
_17
set
_17
statement_timeout = '100ms';
_17
_17
-- authenticated users can run queries that take 5 seconds max
_17
alter
_17
role authenticated
_17
set
_17
statement_timeout = '5s';
_17
_17
-- backend-only users can run queries that take 15 seconds max
_17
alter
_17
role service_role
_17
set
_17
statement_timeout = '15s';

You need to reload PostgREST config cache to apply these changes.


_10
NOTIFY pgrst,
_10
'reload config';

Now, suppose you do an expensive query with the anon role. Like filtering on a big table's unindexed column (this will cause a full table scan):


_10
const { data, error } = await supabase.from('big_table').select().eq('unindexed_column', 'value')

Then, after 5 seconds, the request will be aborted with the response:


_10
{
_10
"hint": null,
_10
"details": null,
_10
"code": "57014",
_10
"message": "canceling statement due to statement timeout"
_10
}

Which is what we wanted. Note that there's already a global statement_timeout set but you can be more fine-grained with this feature. See timeouts for more details.

Statement Cost Limit

With a statement timeout, expensive queries will still get executed for a length of time. They'll consume resources until they’re terminated.

The pg_plan_filter extension (available on the Supabase platform), brings a statement cost limit. This abort queries at the planning phase, before they get executed.

You can use it like:


_22
-- anonymous users can only run cheap queries
_22
ALTER
_22
USER anon
_22
SET
_22
plan_filter.statement_cost_limit = 10000;
_22
_22
-- authenticated users can run more expensive queries
_22
ALTER
_22
USER authenticated
_22
SET
_22
plan_filter.statement_cost_limit = 1e6;
_22
_22
-- backend-only users can run any query
_22
ALTER
_22
USER service_role
_22
SET
_22
plan_filter.statement_cost_limit = 0;
_22
_22
NOTIFY pgrst,
_22
'reload config';
_22
_22
-- reload postgREST config cache to apply changes

Let’s repeat the previous expensive query with the anon role.


_10
const { data, error } = await supabase.from('big_table').select().eq('unindexed_column', 'value')

Then, immediately, the request will be aborted and the response will be:


_10
{
_10
"hint": null,
_10
"details": null,
_10
"code": "54001",
_10
"message": "plan cost limit exceeded"
_10
}

Note that tuning is required to get the cost limit right. You should use the plan_filter.statement_cost_limit with care as it can invalidate legitimate queries.

Configurable Transaction Isolation Level

By default, all queries run in a transaction with the default read committed isolation level.

You can now modify this with the default_transaction_isolation setting.

If you want a function to run with repeatable read isolation level:


_10
create function hello()
_10
returns text as $$
_10
select 'hello';
_10
$$ language sql
_10
set default_transaction_isolation = 'repeatable read';

Or if you want an impersonated role to run its queries with a serializable isolation level:


_10
alter
_10
role service_role
_10
set
_10
default_transaction_isolation = 'serializable';
_10
_10
NOTIFY pgrst,
_10
'reload config';
_10
_10
-- reload postgREST config cache

Note that the default read committed is good enough for almost all use cases. Higher isolation levels incur in overhead as they use more sophisticated locking. They're only needed in special cases.

Bulk insert JSON with default values

A long wanted feature was bulk inserting JSON while considering columns' default values.

Having the following sample table.


_10
create table
_10
foo (
_10
id bigint generated by default as identity primary key,
_10
bar text,
_10
baz int default 100
_10
);

You can now do it like this:


_10
const { error } = await supabase
_10
.from('foo')
_10
.insert([
_10
{ "bar": "val1"
_10
}
_10
, { "bar": "val2"
_10
, "baz": 15
_10
}
_10
], defaultToNull: false)
_10
.select()

And the response will be:


_10
[
_10
{ "id": 1, "bar": "val1", "baz": 100 },
_10
{ "id": 2, "bar": "val2", "baz": 15 }
_10
]

As you can see, id and baz took their default values.

ANY/ALL filter modifiers

As a shortcut to OR filters, you can now use any modifiers on various filters. Take the like filter as an example:


_10
const res = await postgrest
_10
.from('users')
_10
.select()
_10
.likeAnyOf('username', ['%supa%', '%kiwi%'])

This is equivalent to the following in SQL.


_10
select *
_10
from users
_10
where username like ANY('{%supa%,%kiwi%}');

any modifiers are available for the eq,like,ilike,gt,gte,lt,lte,match,imatch filters.

For completeness, the all modifier is also included.

Minimal Breaking Changes from v10

If you only use PostgREST through Supabase client libraries (like supabase-js) then it's safe to upgrade to v11. If you use PostgREST with other HTTP clients (like curl), consider the breaking changes for this version:

  • The Range header is now only considered on GET requests and is ignored for any other method. Previously PostgREST responded with an error but RFC 9110 dictates that we should ignore the Range header instead.
  • RPC requests no longer consider the Prefer: params=multiple-objects header. This header was already deprecated on v10.1.0.

By making use of Logflare, we detected that out of 20 thousands of projects:

  • Only 7 projects used Range for HTTP methods other than GET. In these cases all responses were errors so in fact this breaking change is a fix for those requests.
  • None were using Prefer: params=multiple-objects.

So overall the breaking changes are minimal.

Closing up

There you have it, now you can make your API more secure with role settings and use higher isolation levels without resorting to direct PostgreSQL connections.

PostgREST v11.1 is available for all Supabase projects created after 5 July 2023. Existing projects can upgrade by doing a pause/unpause.

More Postgres resources

Share this article

Build in a weekend, scale to millions