Correlate login and logout times in a table

Question

I have a table LOGINDATA with three columns: USERID, TYPE, and TIMESTAMP.
USERID stores a unique integer that represents someone who uses the system.
TYPE is either 1 or 2. 1 represents a login. 2 represents a logout.
TIMESTAMP is a value that comes from System.currentTimeMillis. It represents the current time in milliseconds since January 1, 1970 (I believe that's the right date).
I'm trying to create a query to link each login time with its corresponding logout time, so that my resultset has three columns:
USERID, LOGINTIME, LOGOUTTIME.
I assume this would involve joining the table to itself, but creating a query to find the closest login and logout times is tricky.
What I have now is:
select li.userid, li.logintime, lo.logouttime from
(select userid, timestamp logintime from logindata where type = 1) li,
(select userid, timestamp logouttime from logindata where type = 2) lo
where li.userid=lo.userid
and li.userid=999999 is the userid of one user.
This will actually end up being more complicated with missing logout times, but this is a start.

Answer 1

The data design makes this a very very very messy (and probably slow) query. Fundamentally, I consider designs like this to be broken because they're so hard to get relevant information out of the raw data; in some cases impossible (see comments below).
The following query works on Oracle database (where TIMESTAMP is a reserved word, so I used EVENTTIME instead) - I tested it with a few rows of data:
select li.userid, li.logintime, min(possible_logout) logouttime
from (select userid || to_char(eventtime, 'yyyymmddhh24miss') session_id, userid, eventtime logintime
      from logindata where type = 1) li,
   (select li.userid || to_char(logintime, 'yyyymmddhh24miss') session_id, li.userid, logintime, logouttime possible_logout
    from (select userid, eventtime logintime from logindata where type = 1) li,
         (select userid, eventtime logouttime from logindata where type = 2) lo
    where logintime < logouttime
      and li.userid = lo.userid(+) ) pl
where li.session_id = pl.session_id (+)
group by li.userid, li.logintime
order by li.logintime;Note that this will return the wrong result if for a particular user_id, you have overlapping logins, like:
time a - login session 1
time b - login session 2
time c - logout session 1 or 2
time d - logout session 2 or session 1
there is simply no way to tell which logout goes with which login and the query will report that both sessions ended at time c, which is absolutely unacceptable if this is to be a security audit trail... The only way to deal with this is to guarantee at the application layer that no user can be logged in twice.
Note also that it breaks by returning extra rows if a user somehow manages to log in twice or more at the same moment, whatever the granularity of your TIMESTAMP is... This is because of the join on the generated SESSION_ID; if you get 2 or more of them the same, you get a cartesian product (2 IDs, 4 rows; 5 IDs, 25 rows, etc...)
It would obviously be a lot cleaner if you stored some sort of unique session_id in your table. Note that the session_ids used for session tracking in J2EE web servers get recycled over time and are therefore unsuitable (by themselves) as unique session ids. And if you store a session id, then you can fix your broken data model and avoid all this mess in the first place by making your table have the columns:
USER_ID, SESSION_ID, LOGIN_TIME, LOGOUT_TIME
Insert on login, update on logout...