They have this response in the other thread:
I would like to explain a recent bug that has been plaguing our system and how it was corrected.
In our system we have a database of all transactions that relate to Satoshidice. We also track their status (pending, confirmed, unknown). We get a handful of transactions that were broadcast but never end up confirming. After 24 to 48 hours we delete them from our system and move on. There was an issue where the status of the transactions was not being recorded correctly. So we would occasionally delete a transaction thinking it was hopeless when it actually confirmed some time ago. Then the outputs spend for that transaction would be marked available and used to pay other transactions. The more I ran the recovery process to clean things up, the more of these ended up getting created.
Now I have a separate safety check that checks an external repository of confirmed transactions before deleting anything. This has been running for about 3 days and seems to work very well.
The other thing that happened that led to a large amount of unconfirming transactions was around Sept 8th we ran out of confirmed funds. We try to always pay winners with confirmed outputs so there is no problem getting them confirmed and they are not dependent on anyone else's transactions. Well, we ran out of confirmed and started paying with pending funds. This of course didn't go well and a bunch of payment transactions had to be deleted and reprocessed.
We of course have alarms on confirmed funds and were notified as it was going down but Erik was traveling and I don't have access to the reserve funds so there was not much I could do, other than later run the recovery and cleanup tasks to get everyone paid.
In our system we have a database of all transactions that relate to Satoshidice. We also track their status (pending, confirmed, unknown). We get a handful of transactions that were broadcast but never end up confirming. After 24 to 48 hours we delete them from our system and move on. There was an issue where the status of the transactions was not being recorded correctly. So we would occasionally delete a transaction thinking it was hopeless when it actually confirmed some time ago. Then the outputs spend for that transaction would be marked available and used to pay other transactions. The more I ran the recovery process to clean things up, the more of these ended up getting created.
Now I have a separate safety check that checks an external repository of confirmed transactions before deleting anything. This has been running for about 3 days and seems to work very well.
The other thing that happened that led to a large amount of unconfirming transactions was around Sept 8th we ran out of confirmed funds. We try to always pay winners with confirmed outputs so there is no problem getting them confirmed and they are not dependent on anyone else's transactions. Well, we ran out of confirmed and started paying with pending funds. This of course didn't go well and a bunch of payment transactions had to be deleted and reprocessed.
We of course have alarms on confirmed funds and were notified as it was going down but Erik was traveling and I don't have access to the reserve funds so there was not much I could do, other than later run the recovery and cleanup tasks to get everyone paid.