In an effort to write less application logic in my services and interfaces, (and out of pure laziness), I am persistently looking for performant strategies to bake access control directly into my schemas. Row Level Security has been around in Postgres for a while now , but we have recently gotten some upgrades to the optimizer which make it worth a deeper investigation.
If you’re not familiar with Row Level Security, the main hypothesis is that we should be able to prevent access to specific rows of data based on a policy. That means our application logic only has to worry about
SELECT * FROM my_table and RLS will handle the
WHERE user_id = my_user_id part automagically.
To put it another way: our queries should only contain the clauses requested by our interfaces and not the filters and conditions demanded by access control in a multi-tenant data store. Take this example:
SELECT id, value FROM items WHERE created_at > now() - interval '1 week' AND items.owner_id = $1
In the above example the
WHERE clause is imposing conditions required for access control, (
items.owner_id ), and requested by an interface, (
created_at ). If our interface grows more complex so could the
WHERE clause and, all of a sudden, separating concerns becomes much harder. Mistakes will be made, and we will keep writing the same annoying access control logic into our queries.
Enter Row Level Security, which provides a space for us to define the conditions demanded by access control separately from the conditions requested by the interface. All of a sudden, our queries look like this:
SELECT id, value FROM items WHERE created_at > now() - interval '1 week'
And, all of a sudden, then can grow as complex as needed without impacting the access control side of the problem.
Since the goal is to produce something usable for real applications, let’s talk requirements:
It should also go without saying that we want our policies to evaluate quickly for large datasets; otherwise, our user experience will suck and no one will like us. I’ll use a couple of three letter initialisms throughout this post:
I will also use
EXPLAIN ANALYZE a lot, which I suggest reading up on if you’re not familiar with it.
Everything to reproduce these tests is available on GitHub , but I’ll just copy the gists on here to keep things simple. Each of the schemas is generated with a dataset of 100 users who can write to 1000 items and read from a subset of the rest.
In general, you can test any of these schemas the following way:
BEGIN; SET LOCAL jwt.claims.roles = ‘c94a743e-8ffd-42dc-a237-b1613e111f53,be95031e-76e0–401a-808c-5c7acde292de’; -- this represents an example query from an interface SELECT count(*) FROM items; COMMIT;
Note that we are using
jwt.claims.roles instead of an existing local variable, such as
current_user , because those behave unexpectedly when querying views and we want this RLS strategy to work for everything, (tables and views). You can also just
SELECT * FROM user_item_stats() , which grabs a random user ID as the
role and executes the previously mentioned transaction.
In a real application, you would swap out the
SELECT statement for whatever interesting query your user wants to run.
This is likely to be the most performant approach to RLS since we won’t need any
SELECT s in our policy; permissions are stored directly on the items. The biggest drawback about storing an ACL in a column is that it will get very wide for items shared with many users/groups, and that may impact performance.
We’ll use the following schema to define our table:
There are two key things which we’ve implemented in the table definition:
ownerscolumn if we needed that as well.
This policy is pretty straightforward: we are just using
&& to check if the claim roles overlap with the ACL. Note that we aren’t doing anything special for public items here; for this schema you’ll have to create a group to represent public instead of a special column as we will do for the next schema. Surprisingly, the planner uses a sequential scan at the deepest node just for adding an
OR items.public = TRUE condition to the
USING clause. That causes a huge performance hit:
Yuck, hopefully performance is better when we avoid the
public logic and handle it like just another group.
As for all of these schemas, I’m testing it using a simple query:
EXPLAIN ANALYZE SELECT count(*) FROM items . Running on my local machine this completes in under 1ms.
Look at all those index scans. Nice.
If we don’t need group-level access we can drop all the sketchy array-splitting logic and just set
jwt.claims.roles to a single ID; regardless, this policy’s performance will not change. If we do want groups, we just need to make sure that
jwt.claims.roles is generated securely in some sort of authorization service. Obviously, a compromised token generator may allow users to inject harmful logic in the local variable.
One of the best things about this implementation is also the worst: the IDs stored in the ACLs are not constrained as foreign keys. This means we can’t add
ON DELETE constraints and the ACL column values may become orphaned. On the other hand, they can reference IDs from multiple tables, like groups and users, which adds a bit of flexibility. Double nice.
As a final test, let’s see what happens when we bump it up to 1000 users and a total of 1 million items:
I won’t make any claims about how performance will change as we add more users, but it safe to say this schema is very fast.
Using rows-in-tables instead of arrays-in-columns feels so much more…normal. The only question: does doing it the “right” way ruin performance? Probably…
For this schema we can take full advantage of a normalized database and foreign key all of the things. To take it one step further, this schema uses an enumerable value for the role of each permission . We could use that
ENUM to add more permission types, (like
owner ), without impacting performance.
The big problem with
REFERENCES is that we can only constrain the permissions table to one other table. Due to this we are forced to store users and groups in one table, which could get confusing. It probably already is confusing, so let me know and I’ll add a comment explaining myself.
Note that we’re also adding some indexes to this permissions table which should help the planner avoid sequential scans.
Here’s a massive policy:
Uh oh, we’ve got a
SELECT in our policy. Those of us who need groups also have an
ANY , it’s not looking good…save us, magical query planner!
Performance is certainly slower at a whopping 22ms for only 100 users, but it is still pretty fast:
The trade-off is that we get to explicitly denote an item as public , and can even add additional permissions roles like
admin which may be important for some use cases. Heck, we could even go as far as a number-based access permission, but who would ever need anything that granular?
Let’s check out what happens when we expand to 1000 users:
We’re still pushing pretty good performance considering how much logic is baked into that policy, but nowhere near as good as the column-based schema.
Moving on, let’s use the same schema as the previous approach but tell the planner what’s up by using a
JOIN s more performantly. Similar to the column-based ACL schema, we’ll have to use a group to implement the public feature.
No policy needed, we can just revoke access to the entire
items table and voilà: the user won’t know the difference. Until they try to
INSERT something into the view, then they’ll know the difference for sure. Also the name might give it away.
ANY isn’t the slowest node in this query, the
JOIN forced a sequential scan for
item_id = items.id . Maybe I’m getting tired or maybe it’s impossible, either way I’m disappointed by the performance of this schema.
Another huge drawback: this view is not automatically updatable, which means users will still need to access the underlying table. They can do that either through stored procedures or the same
WITH CHECK -type policy that we used in the table implementation. While this may have been a more performant approach pre-pg@10, it definitely isn’t holding up well now.
That said, if you don’t need groups, dropping the
ANY clause does yield significant improvements:
Both of the RLS schemas performed well and may be valuable tools for different use cases. The table-based approach avoids modifying existing tables, while the column-based approach if blazingly fast. In my mind, the only thing to do next is trying building something with both of them and see what pain points emerge.
If you have any optimizations for these schemas drop them in the comments, it would be great to push them to their limits. Additionally, if you’ve found alternative strategies which offer great performance via sane schemas I’d love to hear about them.