On Tue, 2003-02-11 at 12:10, Shane Hathaway wrote:
I added a test to testZODB.py on a new branch (shane-conflict-handling-branch) that exercises the conflict handling bug. The test currently fails. It might be simpler to go with Toby's implementation for now: add a "veto" object to the transaction that refuses any attempt to commit. But maybe your transaction states are better. Let me know what you want to do.
I'd like to do the transaction states, because it would keep the code in zodb3 and zodb4 similar. Unless there's a reason to think there are problems with the transaction state approach. I didn't look carefully at the test, but if I remember the discussion last time around, the problem is with read conflicts caught outside of 2PC. In that case, we either need to mark the connection so that it votes no when it gets to prepare() or we need to veto() method. I'd prefer the vote-no-in-prepare because it keeps the API smaller, but veto() isn't so bad; maybe it's better to stop the transaction quickly. We have a similar need when things go wrong inside the persistent machinery. I can't recall the details, but I think there are comments in the code about what happens when loading an object fails -- some state transition fails at any rate. In that case, the transaction really needs to be marked as failed and further actions prevented. Jeremy