Hi Dormando,
Normally in our application memcache is used for the purpose of caching
database values but someone from the implementation team decided to use
memcache for locking which lead us to this code that works 99% only. I am
just trying to make it better especially since we know that race conditions
exists.
In this case, I am trying to address a screen where we deal with inbound
documents (read from a file system) that needs to be moved to our DB
depending on the context. We want to avoid multiple users working on the
same document as much as possible. If a document is in his cart, we
consider that he is working on it. Its a b-2-b application so that is
different from the normal b-2-c applications where cart means something
else. We help other companies to use our tool to enchance their
productivity.
Locks are by default always tied up to the user session. So when
administrator userlogs out, his cache is always cleared or if his session
expires. A time-out is always implict for every user abd by default it is
10 minutes. He can also remove the locks if he wants. Further on a daily
basis the memcache is cleared as well. [We store a key like filename: user
when a user locks a file which is what we have been discussing most of the
same. We also append a user key which has a list like [filename,filename]
to ensure that both lookups are possible and clear outs are easy when the
user logs out etc. That reminds be that I probably need to use cas for the
appending because of race conditions.]
Overall I felt that the scenario is some what different here and thought of
clarifying everything before I can propose something. I have lots of time
on this because it is something I am doing voluntarily. It has not been my
intention to ignore your advise but I wanted to make sure the use case is
really clear but it seemed like we are doing it somewhat differently. It is
my mistake that I didn't mention that session time out and daily cache
clearing happens which automatically takes care of removing locks if he
doesn't voluntarily give up the lock. As mentioned if he moves out of the
Document screen, we clear it as well.
Based on this please do let me know your thoughts. Also I feel one flaw so
far in my logic is that I need to check if memcache is really up otherwise
just proceed as usual. We will get more errors but that is better than
stopping productivity.
Answering some of your questions:
What if a user adds something to their cart and never comes back? It gets
expired if he navigates out of the Documents screen in our SPA. It also
gets expired when he times out or if he manually remove the lock. It will
happen anyway at the end of the day if neither of these happens.
Why don't we use a database? I think we should use database but probably it
was not considered due to dynamic nature of this table, but I can't answer
that since it their decision so perhaps we just forward this to them.
I hope I am not taking out your time. There is no hurry on this like I said
I voluntarily trying to make it a bit better.
Thanks,
Nishant
On Wednesday, June 8, 2016 at 11:58:00 PM UTC+5:30, Dormando wrote:
>
>
>
> On Tue, 7 Jun 2016, Nishant Varma wrote:
>
> > There was too much noise removed all those posts :-)
> > key = filename
> > item = memcli:get(key)
> > if (! defined item) {
> > if (memcli:add(key . "_lock", lock_timeout_time,
> my_admin_username)) {
> > [etc]
> > } else {
> > # lost the race, handle how you want
> > }
> > } else if (item.value == my_admin_username) {
> > # good to go for that future request
> > }
> >
> >
> > In my scenario files are displayed from a certain folder (queue) in an
> application to many users at the same time. Each users has a cart where
> they can add files they are
> > working upon. They are going to remove it from the cart or process it
> from the cart.
> >
> > In the example pattern you have given,
> >
> > 1) The locking and the action are both happening together. However, we
> want lock to move the item to the cart when the user clicks on the file. If
> items are there in his cart,
> > he can do anything he wants. Lock out time is also not needed, user will
> remove it from the cart. [There is a flaw here what if someone manipulates
> dom and add stuff to cart -
> > but we won't worry about that scenario - your example actually takes
> care of that since validation and action if combined]
>
> What if a user adds something to their cart and never comes back? After
> some amount of time you'd want to re-validate what's in their cart. ie;
> fetch all of the values again, and do the add-dance if the gets don't say
> they still own it. That's why the timeout is useful, even if it's very
> long (like 8 hours, or days even).
>
> If they lose their cart and log in again, do they lose all of their locks?
> Is the cart stored in a session in a database somewhere?
>
> > 2) There is a different key called filename + lock which feels a bit
> redundant in my case because I can achieve everything I want by entering
> the key as filename and like you
> > said username as the value.
> >
> > I have loads of time to get this code in so I wanted to review this in
> detail before suggesting something.
>
> ok. I'm a bit confused because you've been swapping between use cases? at
> first it was some kind of admin deal, now locking files, etc?
>
> > It has one major flaw, memcache.add can also be False because server is
> down. I will create a blocker because now users can't add anything -
> reducing with productivity at the
> > cost of stopping errors. So I also need to check if memcache client is
> up and running, else don't worry about locks. That code is still not added.
>
> I've said it a few times: the ghetto locking is for *advisory locks only*,
> if you're gating real functionality on it you are severely, severely doing
> it wrong. This must be accelerating something which is already in a
> database. The system is designed to be lossy and tolerant to failure, and
> it achieves that by being a layer on top of something else, conceptually.
> Doing it this way guarantees pain forever: you will never want to restart
> memcached, you will never want to upgrade it, you will never want to tune
> it, because restarting it will break all of your active locks.
>
> Is your database so slow that you can't note locks in it? Do you not have
> a database and are using memcached as a layer in front of a file store?
>
> It's very hard to help you. I want to, but you've been selectively
> ignoring advice and it's hard to recommend using this method unless I
> understand your use case extremely well: as I was asking above, what
> happens when carts go away, what is the usage of a cart, how is a user
> actually interacting with this thing, etc. Your own code below says it
> breaks if memcached is down, does it handle a restart?
>
> sorry. wish I could do more.
>
> >
> > def addToCart(filename, username):
> > ableToLock = memcache.add(filename, username)
> > if ableToLock:
> > # ableToLock can happen if the file is still present
> > # or if it was already processed.
> > if os.file.ispath(filename):
> > # I have a lock and file exists. Think of Cart as a
> > # JS Object from where you can pick items to "process".
> > return "Added To Cart"
> > else:
> > # I have a lock but looks like file was processed already.
> > # So removing the residual "key" created.
> > memcache.delete(filename)
> > return "Processed by another user."
> > else:
> > # "add" can also fail if remote server is down,
> > # but now we are not handling that now. It will
> > # block the user's ability to process anything.
> > user = memcache.get(filename)
> > if user and os.file.ispath(filename):
> > # I try my best to show the user processing it.
> > print "Being processed by %s" % user
> > else:
> > # But lost the race to find that.
> > print "Processed by another user."
> >
> >
> > def process(source):
> > shutil.move(source, destination)
> > # filename => source
> > memcache.delete(filename)
> >
> >
> >
> >
> > On Sunday, June 5, 2016 at 1:30:37 AM UTC+5:30, Dormando wrote:
> > The pattern is identical between the one in the wiki and yours.
> Simply
> > move the delete of the key until you're done using the lock, which
> would
> > be in a separate request.
> >
> > In your case, you would probably set the contents of the key to be
> the
> > name of the user who has it locked.
> >
> > In the original pseudocode:
> > key = "expensive_frontpage_item"
> > item = memcli:get(key)
> > if (! defined item) {
> > # Oh crap, we have to recache it!
> > # Give us 60 seconds to recache the item.
> > if (memcli:add(key . "_lock", 60)) {
> > item = fetch_expensive_thing_from_database
> > memcli:add(key, item, 86400)
> > memcli:delete(key . "_lock")
> > } else {
> > # Lost the race. We can do any number of things:
> > # - short sleep, then re-fetch.
> > # - try the above a few times, then slow-fetch and return
> the item
> > # - show the user a page without this expensive content
> > # - show some less expensive content
> > # - throw an error
> > }
> > }
> > return item
> >
> > In yours:
> > key = filename
> > item = memcli:get(key)
> > if (! defined item) {
> > if (memcli:add(key . "_lock", lock_timeout_time,
> my_admin_username)) {
> > [etc]
> > } else {
> > # lost the race, handle how you want
> > }
> > } else if (item.value == my_admin_username) {
> > # good to go for that future request
> > }
> >
> > Then when you're done holding the lock, delete the key.
> >
> > On Sat, 4 Jun 2016, Nishant Varma wrote:
> >
> > > I am reading
> https://github.com/memcached/memcached/wiki/ProgrammingTricks#ghetto-central-locking,
>
> it seems to deal with a slightly different lock scenario of
> > getting some
> > > expensive item from Database to avoid "Stampeding"
> > > In my case its slightly different lock that I need. I show
> regular files from a folder in a web application to many users. So, to
> "lock" a file using
> > memcache isn't this
> > > simple API sufficient or I still need that pattern :-)?
> > > def access(filename):
> > > if memcache.add(filename, timestamp):
> > > return "Access Granted. Lock Obtained" # Normally this
> results in checking HTML checkbox against the filename so User can do
> actions with that/
> > > else:
> > > return "Access Denied" # Normally this leads to an alert
> saying that someone else is working on this.
> > >
> > > Isn't this simple API using add good enough in my case? I am
> sorry if I am repeating this, but I could not really relate the "fetching
> expensive item from
> > Database" to my
> > > scenario which is why I even wrote a simple script to test the
> validity of the claim etc.
> > >
> > > Can you please let me know?
> > >
> > >
> > > On Saturday, June 4, 2016 at 6:42:35 PM UTC+5:30, Nishant Varma
> wrote:
> > > Excellent I rely on you. I guess this is the reason you
> say I am over-engineering this problem. Makes sense :-) I will again check
> the link you gave me. I
> > will go
> > > through the documentation this weekend.
> > >
> > > On Saturday, June 4, 2016 at 1:33:04 PM UTC+5:30, Dormando
> wrote:
> > > Hey,
> > >
> > > You really don't need to test this: I'm telling you
> flatly, as an author
> > > of this software and all of the documentation for
> it, that you should
> > > absolutely not rely on that pattern. I'm trying to
> save you some time.
> > >
> > > The pattern that is slightly better is written
> explicitly in pseudocode in
> > > the link I gave you several times in the issue.
> Please use it.
> > >
> > > Thanks,
> > > -Dormando
> > >
> > > On Fri, 3 Jun 2016, Nishant Varma wrote:
> > >
> > > > Can anyone help me peer review this script
> https://gist.github.com/varmanishant/0129286d41038cc21471652a6460a5ff
> that demonstrate potential
> > problems
> > > with get set if it is used
> > > > to implement distributed locking. I was suggested
> to modify from get set to add in this thread
> https://github.com/memcached/memcached/issues/163.
> > > However I wanted a small
> > > > simulation to demonstrate this.
> > > >
> > > > --
> > > >
> > > > ---
> > > > You received this message because you are
> subscribed to the Google Groups "memcached" group.
> > > > To unsubscribe from this group and stop receiving
> emails from it, send an email to [email protected].
> > > > For more options, visit
> https://groups.google.com/d/optout.
> > > >
> > > >
> > >
> > >
> > > This e-mail message (including any attachments) may contain
> information that is confidential, protected by the attorney-client or other
> applicable privileges, or
> > otherwise
> > > comprising non-public information. This message is intended to
> be conveyed only to the designated recipient(s). If you have any reason to
> believe you are not an
> > intended
> > > recipient of this message, please notify the sender by replying
> to this message and then deleting it from your system. Any use,
> dissemination, distribution, or
> > reproduction of
> > > this message by unintended recipients is not authorized and may
> be unlawful.
> > >
> > > --
> > >
> > > ---
> > > You received this message because you are subscribed to the
> Google Groups "memcached" group.
> > > To unsubscribe from this group and stop receiving emails from
> it, send an email to [email protected].
> > > For more options, visit https://groups.google.com/d/optout.
> > >
> > >
> >
> >
> > This e-mail message (including any attachments) may contain information
> that is confidential, protected by the attorney-client or other applicable
> privileges, or otherwise
> > comprising non-public information. This message is intended to be
> conveyed only to the designated recipient(s). If you have any reason to
> believe you are not an intended
> > recipient of this message, please notify the sender by replying to this
> message and then deleting it from your system. Any use, dissemination,
> distribution, or reproduction of
> > this message by unintended recipients is not authorized and may be
> unlawful.
> >
> > --
> >
> > ---
> > You received this message because you are subscribed to the Google
> Groups "memcached" group.
> > To unsubscribe from this group and stop receiving emails from it, send
> an email to [email protected] <javascript:>.
> > For more options, visit https://groups.google.com/d/optout.
> >
> >
--
This e-mail message (including any attachments) may contain information
that is confidential, protected by the attorney-client or other applicable
privileges, or otherwise comprising non-public information. This message is
intended to be conveyed only to the designated recipient(s). If you have
any reason to believe you are not an intended recipient of this message,
please notify the sender by replying to this message and then deleting it
from your system. Any use, dissemination, distribution, or reproduction of
this message by unintended recipients is not authorized and may be unlawful.
--
---
You received this message because you are subscribed to the Google Groups
"memcached" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
