home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.databases
- Path: sparky!uunet!ulowell!news.bbn.com!usc!sdd.hp.com!hpscdc!cupnews0.cup.hp.com!dhepner
- From: dhepner@cup.hp.com (Dan Hepner)
- Subject: Re: Hot Standby DBMS's
- Sender: news@cupnews0.cup.hp.com
- Message-ID: <BtruJB.Lrn@cup.hp.com>
- Date: Sun, 30 Aug 1992 00:54:47 GMT
- References: <CHRIS.92Aug28101634@yen.imsi.com>
- Organization: Hewlett-Packard
- X-Newsreader: Tin 1.1scd1 PL4
- Lines: 36
-
- From: chris@yen.imsi.com (Chris Payne)
-
- > If this advice "credit must always precede debit" is not followed you
- > might get deadlocks even without replication. However, this is
- > insufficient to prevent the other kind of out-of-order deadlocks.
-
- >Here's the idea: serialize & fork the db request stream. Each hot
- >standby database will receive the same transactions in the same order.
- >Other receiver(s) can take the same stream and keep it for posterity.
- [but ...]
-
- We'll agree that your solution both solved the out-of-order lock problem,
- but left an unusable solution due to performance considerations.
- The problem is that you went way too far. One need not serialize
- the entire input stream.
-
- > >If you want to catch-up a server which is down periodically, keep a
- > >copy of the request stream.
- >
- > How would you guarantee that transactions will be executed in the same
- > serial order as was implied by the online system? As has been pointed
- > out, such a guarantee is required to keep the copies synchronized.
- >
- >Unless you're worried about the non-serial nature of the server
- >itself, this is trivial, right?
- >/c (chris@imsi.com)
-
- It's not trivial; while the transactions were of course executed in
- parallel, there is an "as if serialized" ordering which was achieved.
- We cannot know that ordering until after the fact, and even then the only
- way we are likely to be able to figure it out is by examining proprietary
- log records. However, if we expect to keep two replicates perfectly
- synchronized, this ordering must be exact on all replicates. This
- generally precludes a solution based on a queued request stream.
-
- Dan Hepner
-