对数据库（Oracle 或 ANSI SQL）中具有选定 id 的行进行比较和交换操作？

Question

I want to implement failsafe persistent synchronization in application based on database (Oracle in my case, but like to see ANSI SQL solution).

I work with tasks which can be run in different threads, applications or servers.

Each type of task (I differ them by ID ) may not run in concurrent - that is why I need synchronization. All have access to DB so it is good place for synchronization!

Each thread/application/server can fail, so I need a way to remove acquired lock from ID after timeout.

The first that come to mind is to use table with:

ID
STATE
TS

fields. All I need is atomic operations which:

• try to change STATE value from completed to executing (to synchronize) and set TS to current time. Return false if STATE is not completed .
• try to change STATE value from executing / recovering to recovering if sysdate - TS > delay and set TS to current time (to be failsafe). Else return false.

SQL update statement mostly make that I want:

update TASK set STATE = 'executing', TS = sysdate
  where ID = :id and STATE = 'completed'

and:

update TASK set STATE = 'recovering', TS = sysdate
  where ID = :id and STATE in ('executing', 'recovering')
    and sysdate - TS > :delay

Only one issue that I see - how to get know (from Java application through JDBC) if update actually performed or not (in order to be true compare-and-swap operation )? May be by getting updated row number (is this info available through JDBC)?

Is I am correct with my assumption that update is atomic for where condition?

Is there another way to implement failsafe persistent synchronization in application based on database?

PS My question is differ from:

• atomic compare and swap in a database (it have 2 select and update statement, I end with one).

Answer 1

This easiest way might be to create and run a stored procedure which could then return what the result of the change was.

Edit (expanded answer): Sorry I thought you were looking for a way to get the answer back. I would do something like the following (although I haven't got an oracle instance at home to test it on, sorry)

function GetLockForRun(p_id task.id%TYPE)
  return VARCHAR2
declare
 l_result VARCHAR2;
begin
 select STATE, TS
   into l_state, l_ts
   from Task
   where ID = p_id;
     for update; --this locks the row until you commit 

 if (state == 'completed' or sysdate - TS > delay)
   update TASK 
      set STATE = 'executing', TS = sysdate
    where ID = p_id;
   l_result := "OK";
 else 
   l_result := "Do not run";
 end if
 commit; --release for update lock
 return l_result;
end;

This function locks the row for the length of its execution so no other process can edit it, this means only one process can run this at a time. Since this procedure will run in the database it will finish so you don't have to worry if the java process dies. I guess the only downside is that this is blocking but only for the run of this function which is short. If you really can't have that then you could try making the select for update not wait; off the top of my head I can not remember how to do this.