simultaneous multiple requests to very simple database

[Eric S. Johansson]

I have an application where I need a very simple database, effectively a
very large dictionary. The very large dictionary must be accessed from
multiple processes simultaneously. I need to be able to lock records
within the very large dictionary when records are written to. Estimated
number of records will be in the ballpark of 50,000 to 100,000 in his
early phase and 10 times that in the future. Each record will run about
100 to 150 bytes.

speed is not a huge concern although I must complete processing in less
than 90 seconds. The longer the delay however the greater number of
processes must be running parallel in order to keep the throughput up.
It's the usual trade-off we have all come to know and love.

it is not necessary for the dictionary to persist beyond the life of the
parent process although I have another project coming up in which this
would be a good idea.

at this point, I know they will be some kind souls suggesting various
SQL solutions. While I appreciate the idea, unfortunately I do not have
time to puzzle out yet another component. Someday I will figure it out
because I really liked what I see with SQL lite but unfortunately, today
is not that day (unless they will give me their work, home and cell
phone numbers so I can call when I am stuck. ;-)

So the solutions that come to mind are some form of dictionary in shared
memory with locking semaphore scoreboard or a multithreaded process
containing a single database (Python native dictionary, metakit, gdbm??)
and have all of my processes speak to it using xmlrpc which leaves me
with the question of how to make a multithreaded server using stock xmlrpc.

so feedback and pointers to information would be most welcome. I'm
still exploring the idea so I am open to any and all suggestions (except
maybe SQL :-)

---eric

[Thomas Bartkus]

"Eric S. Johansson" <es*@harvee.org> wrote in message
news:ma**************************************@pyth on.org...
<snip>

at this point, I know they will be some kind souls suggesting various
SQL solutions. While I appreciate the idea, unfortunately I do not have
time to puzzle out yet another component. Someday I will figure it out
because I really liked what I see with SQL lite but unfortunately, today
is not that day (unless they will give me their work, home and cell
phone numbers so I can call when I am stuck. ;-)

<snip>

Forgive me if this reply sounds a bit glib. But I do mean it without malice.

Do you seriously expect to write your own (database) solution and that this
will save you time and effort over learning an existing (SQL) solution?

Because -
If you are seeking to "save time" on "puzzles", you are certainly going
about it the wrong way.

Best of luck
Thomas Bartkus

[Eric S. Johansson]

Thomas Bartkus wrote:

"Eric S. Johansson" <es*@harvee.org> wrote in message
news:ma**************************************@pyth on.org...
<snip>

at this point, I know they will be some kind souls suggesting various
SQL solutions. While I appreciate the idea, unfortunately I do not have
time to puzzle out yet another component. Someday I will figure it out
because I really liked what I see with SQL lite but unfortunately, today
is not that day (unless they will give me their work, home and cell
phone numbers so I can call when I am stuck. ;-)
<snip>

Forgive me if this reply sounds a bit glib. But I do mean it without malice.

understood and taken in that spirit.
Do you seriously expect to write your own (database) solution and that this
will save you time and effort over learning an existing (SQL) solution?

Because -
If you are seeking to "save time" on "puzzles", you are certainly going
about it the wrong way.

one thing I learned a long time ago was to respect the nagging voice in
the back of my head that says "there is something wrong". Right now
with databases, that voice is not nagging but screaming. So I made my
query to try and prove that intuition wrong. So far, that has not happened.

When I look at databases, I see a bunch of very good solutions that are
either overly complex or heavyweight on one hand and very nice and
simple but unable to deal with concurrency on the other. two sets of
point solutions that try to stretch themselves and the developers to fit
other application contexts.

99.9 percent of what I do (and I suspect this could be true for others)
could be satisfied by a slightly enhanced super dictionary with a record
level locking. but, the database world does not fit this model. It has
a great deal more complication then what is frequently necessary.

If I ever find the time, I will try to build such a beast probably
around Metakit. The only reason for reluctance is that I have spent too
many hours tracking down concurrency problems at the OS level way to
many years ago and so I do not create multithreaded applications lightly.

so in conclusion, my only reason for querying was to see if I was
missing a solution. So far, I have not found any work using because
they add orders of magnitude more complexity than simple dbm with file
locking. Obviously, the simple solution has horrible performance right
now I need simplicity implementation.

thanks for your commentary.

---eric

[Ricardo Bugalho]

On Tue, 18 Jan 2005 17:33:26 -0500, Eric S. Johansson wrote:

When I look at databases, I see a bunch of very good solutions that are
either overly complex or heavyweight on one hand and very nice and simple
but unable to deal with concurrency on the other. two sets of point
solutions that try to stretch themselves and the developers to fit other
application contexts.

Have you considerded SQLite/pySQLite ?

--
Ricardo

[Stephen Thorne]

On Tue, 18 Jan 2005 17:33:26 -0500, Eric S. Johansson <es*@harvee.org> wrote:

so in conclusion, my only reason for querying was to see if I was
missing a solution. So far, I have not found any work using because
they add orders of magnitude more complexity than simple dbm with file
locking. Obviously, the simple solution has horrible performance right
now I need simplicity implementation.

thanks for your commentary.

Maybe you can just get the best of both worlds.

Have a look at SQLObject. You can ignore the fact that underneath the
SQLObject there's a postgres (or mysql, or whatever) database, and get
OO based persistance.

SQLObject is crippled in that there are degrees of freedom that SQL
gives you that SQLObject takes away/makes hard to use, but what you're
trying to do, and what most people actually do with databases, can be
easily wrapped around with a simple, pythonic wrapper.

It even has a .createTable() function for those times when you don't
even want to log into the database.

Regards,
Stephen Thorne.

[Eric S. Johansson]

Ricardo Bugalho wrote:

On Tue, 18 Jan 2005 17:33:26 -0500, Eric S. Johansson wrote:

When I look at databases, I see a bunch of very good solutions that are
either overly complex or heavyweight on one hand and very nice and simple
but unable to deal with concurrency on the other. two sets of point
solutions that try to stretch themselves and the developers to fit other
application contexts.

Have you considerded SQLite/pySQLite ?

yep and apparently it won't work

http://www.sqlite.org/faq.html#q7

if I had record level locking, the code would do a very common pattern like:

if record present:
Lock record
modify record
release lock
else:
create record atomically (actual method TBB)

if I read their opinion correctly, the SQL lite folks are wrong in that
only the applications need massive concurrency. Small applications need
significant to massive concurrency for very tiny windows on very little
data.

but I do appreciate the pointer.

[Thomas Bartkus]

"Eric S. Johansson" <es*@harvee.org> wrote in message
news:ma**************************************@pyth on.org...
<snip>

99.9 percent of what I do (and I suspect this could be true for others)
could be satisfied by a slightly enhanced super dictionary with a record
level locking.
BUT - Did you not mention! : Estimated number of records will be in the ballpark of 50,000 to 100,000 in his early phase and 10 times that in the future. Each record will run about 100 to 150 bytes. ..
And The very large dictionary must be accessed from
multiple processes simultaneously
And I need to be able to lock records
within the very large dictionary when records are written to
And although I must complete processing in less than 90 seconds.

And - the hole in the bottom of the hull -
all of the above using "a slightly enhanced super dictionary".

*Super* dictionary??? *Slightly* enhanced???
Have you attempted any feasability tests? Are you running a Cray?

There are many database systems available, and Python (probably) has free
bindings to every one of them. Whichever one might choose, it would add
simplicity, not complexity to what you are attempting. The problems you
mention are precisely those that databases are meant to solve. The only
tough (impossible?) requirement you have is that you don't want to use one.

When you write that "super dictionary", be sure to post code!
I could use one of those myself.
Thomas Bartkus

[Eric S. Johansson]

Thomas Bartkus wrote:

When you write that "super dictionary", be sure to post code!
I could use one of those myself.

hmmm it looks like you have just flung down the gauntlet of "put up or
quityerwhinging". I need to get the crude implementation done first but
I think I can do it if I can find a good XMLRPC multithreading framework.

---eric

[Jeremy Sanders]

On Tue, 18 Jan 2005 11:26:46 -0500, Eric S. Johansson wrote:

So the solutions that come to mind are some form of dictionary in shared
memory with locking semaphore scoreboard or a multithreaded process
containing a single database (Python native dictionary, metakit, gdbm??)
and have all of my processes speak to it using xmlrpc which leaves me
with the question of how to make a multithreaded server using stock
xmlrpc.

Another solution might be to store the records as files in a directory,
and use file locking to control access to the files (careful over NFS!).

You might also consider berkeley db, which is a simple database to add to
an application, (and which I believe supports locks), but I must admit I'm
not a fan of the library.

I assume that the bottleneck is processing the records, otherwise this all
seems a bit academic.

Jeremy

[Tom Loredo]

Just learned of this today, so I don't know enough details to judge
its suitability for you:

Durus
http://www.mems-exchange.org/software/durus/

It does not do locking, but alleges to be compact and easy to
understand, so perhaps you could modify it to meet your needs,
or find some other way to handle that requirement.

-Tom

--

To respond by email, replace "somewhere" with "astro" in the
return address.

[phr]

I agree with you, there's a crying need for something like that and
there's no single "one obvious way to do it" answer.

Have you looked at bsddb? See also www.sleepycat.com.