\df *crypt
in psql reveals the argument types of the pgcrypto encrypt
and decrypt
functions (as do the PgCrypto docs):
List of functions
Schema | Name | Result data type | Argument data types | Type
--------+-----------------+------------------+--------------------------+--------
...
public | decrypt | bytea | bytea, bytea, text | normal
public | encrypt | bytea | bytea, bytea, text | normal
...
so both the encrypt
and decrypt
functions expect the key to be bytea
. As per the error message, "you might need to add explicit type casts".
However, it works fine here on Pg 9.1, so I suspect there's more to it than you've shown. Perhaps you have another function also named encrypt
with three arguments?
Here's how it works on a clean Pg 9.1:
regress=# create table demo(pw bytea);
CREATE TABLE
regress=# insert into demo(pw) values ( encrypt( 'data', 'key', 'aes') );
INSERT 0 1
regress=# select decrypt(pw, 'key', 'aes') FROM demo;
decrypt
------------
\x64617461
(1 row)
regress=# select convert_from(decrypt(pw, 'key', 'aes'), 'utf-8') FROM demo;
convert_from
--------------
data
(1 row)
Awooga! Awooga! Key exposure risk, extreme admin caution required!
BTW, please think carefully about whether PgCrypto is really the right choice. Keys in your queries can be revealed in pg_stat_activity
and the system logs via log_statement
or via crypto statements that fail with an error. IMO it's frequently better to do crypto in the application.
Witness this session, with client_min_messages
enabled so you can see what'd appear in the logs:
regress# SET client_min_messages = 'DEBUG'; SET log_statement = 'all';
regress=# select decrypt(pw, 'key', 'aes') from demo;
LOG: statement: select decrypt(pw, 'key', 'aes') from demo;
LOG: duration: 0.710 ms
decrypt
------------
\x64617461
(1 row)
Whoops, key possibly exposed in the logs if log_min_messages
is low enough. It's now on the server's storage, along with the encrypted data. Fail. Same issue without log_statement
if an error occurs to cause the statement to get logged, or possibly if auto_explain
is enabled.
Exposure via pg_stat_activity
is also possible.. Open two sessions, and:
- S1:
BEGIN;
- S1:
LOCK TABLE demo;
- S2:
select decrypt(pw, 'key', 'aes') from demo;
- S1:
select * from pg_stat_activity where current_query ILIKE '%decrypt%' AND procpid <> pg_backend_pid();
Whoops! There goes the key again. It can be reproduced without the LOCK TABLE
by an unprivileged attacker, it's just harder to time it right. The attack via pg_stat_activity
can be avoided by revoking access to pg_stat_activity
from public
, but it just goes to show that it might not be best to send your key to the DB unless you know your app is the only thing ever accessing it. Even then, I don't like to.
If it's passwords, should you store them at all?
Furthermore, if you're storing passwords, don't two-way encrypt them; if at all possible salt passwords then hash them and store the result. You usually don't need to be able to recover the password cleartext, only confirm that the stored hash matches the password the user sends you to log in when it's hashed with the same salt.
If it's auth, let someone else do it for you
Even better, don't store the password at all, authenticate against LDAP, SASL, Active Directory, an OAuth or OpenID provider, or some other external system that's already designed and working.
Resources
and lots more.
Best Answer
The difference is this:
In the first two (working) examples,
'a'
is just an untyped string literal - which is strictly different fromvarchar
. The conversion to your customenum
type is provided by basic input/output functions of the enum type. (There is no explicit entry in the system catalogpg_cast
.) Postgres determines the best match for a data type as the user did not provide one. Your custom cast is not even needed. Neither for type resolution nor for operator resolution.In the following examples, you pass a typed value (
varchar
). Now, your custom cast may be of use. For example, this wouldn't work without your custom cast:Your examples are not plain assignments, but expressions with the operator
=
. We are in the realm of Operator Type Resolution. Item 2a states:That would cover
my_enum_column = 'foo'
- with untyped literal.But there is nothing that can be stretched to cover
my_enum_column = 'foo'::varchar
. Seems like the fact that only a generic=
operator for left and right operandanyenum
is available is not enough to clue in operator resolution:You would have to register a more explicit operator for this. But I wouldn't do that. I would do things like (also note various orthogonal syntax fixes):
db<>fiddle here
Or just add explicit casts. That's simplest and safest. The manual warns (on the same page you quoted):
There is more - recommended reading.
If you must create that custom cast (or even add an operator), consider using
text
instead ofvarchar
, which is the "preferred type" among string types and hence more robust against possible corner-case issues.