I posed your question to a colleague of mine, because this problem is way beyond my puny knowledge. Anyway, his response, for what it's worth:

>>>>>
My first instinct is to call this a Supremely Bad Idea.  There WILL be
conflicts.  You can bet on it.  Best to have a single "maintenance"
process on a single host that fails over as needed.  Alternatively, rely
on shared/exclusive table-level locking which is extremely efficient,
widely supported, and does not suffer from the issues that row-level
locking is famous for.

If the random number generator is still insisted upon, there are
standard solutions for seeding it.  The java.security.SecureRandom class
uses host and network performance and latency metrics to seed itself.
While that might not be quite as well-seeded as a "pot of boiling
water", it's pretty darn close.

If that's not sufficient and you're running on a UNIX-style system
(Linux, BSD, Solaris, etc), you can read a block from /dev/random and
seed your generator with that.  It's a system-level random state that
survives reboot.

Finally, you can seed it with a UUID.  Most containers these days
provide some sort of UUID service, and if not then there are example
UUID generators all over the net.
<<<<<

<original uncommented post left below for reference:>

Frank W. Zammetti wrote the following on 3/18/2005 10:35 AM:
Oh boy, I got a good one!  It's only related to Struts in that the
application in question is Struts-based, so I hope no one minds a semi-OT
question...

Here's the situation... An app I wrote has a daemon thread that is spawned
at startup (from a Struts plugin) that does periodic background processing
tasks.  This works great, never had a bit of trouble.

Now though, the app is moving from a single server to a clusted environment.

So, what's going to happen is that each server in the cluster will have
its own instance of the thread running on it.  Not a huge problem except
that I have to be sure only one instance of the thread (i.e., one server
in the cluster) is executing concurrently.

The easy solution is just a database table that is checked when the thread
wakes up.  If there is no entry in it, then there is no other instance
running, so it can write an entry to the table and go off and do its
thing.

I want to be extremely certain that no issues arise in terms of one
instance of the thread reading from the database while another instance is
writing, etc.  So, aside from transactional database calls and row-level
locking, I want to do one more thing: I want the thread to sleep a random
number of seconds (1-300) at startup.  This will ensure that, all the
database locking and such aside, the threads should all be offset from one
another in terms of when they run.

So, I need a random number generated when the thread starts up.  As we all
know though, random number generation on most computers that don't have
something like a Brownian motion sensor attached stuck in a cup of boiling
coffee can't generate truly random numbers.  So, in theory, what could
happen is that if all the servers in the cluster come up at the same time,
the threads could wind up running at the same time regardless of the
random sleep at the start!  It might never happen in reality, small
fluctuations would probably offset them anyway, but I want to be more
certain than that.

So now we're at the crux of the problem...

I can't just seed the random number generator with the current time
because it concievably might not be random enough.  So, I thought I could
just tally up the octets of the server's IP address and add that to the
current time.  Then the seed on each server should be different enough.

But, there doesn't appear to be any way to get the server IP address
independant of a request, so I can't get at it in my plugin.  Anyone know
differently?

Assuming that is the case, can anyone think of any other way to seed the
generator that would ensure a different value on different machines in the
cluster?  There are some options like encoding the individual server names
in my app's config file with a different seed value for each, but that
makes maintenance a pain if a new server is added or one removed or
addresses simply changed.

Any ideas?  Thanks!



--
Rick

---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]



Reply via email to