On 11/26/18 1:23 PM, Phil Steitz wrote:
On 11/26/18 8:29 AM, Phil Steitz wrote:
On 11/26/18 6:19 AM, Mark Struberg wrote:
Hi Phil!
Let me start by repeating that other than trying to help diagnose
bugs and answer user questions, I don't really work on [pool] any
more, so I don't really have any standing here. You are the RM,
so it is completely up to you and the active Commons committers
to decide what to do with the code. So fine to ignore my comments.
I think that the right answer here is WONT_FIX. That sounds
harsh, but it seems to me
that the obvious solution here is for the user to set maxIdle
to at least 1.
I thought about that as well!
Setting maxIdle to 1 will make it less likely but a deadlock
will STILL happen.
I am not sure this is right. If maxIdle >0, the fix for POOL-240
(a similar liveness issue) should kick in and work. Each time a
validation or passivation failure occurs on return, there is a
call to ensureIdle that will create a new instance and add it to
the pool if there is capacity to do so.
So I fear we really need to tackle this. Stackoverflow and our
own bug tracker is full of such reports :(
I see one additional actual report on GKOP (the other linked
issue, which should be similarly patched if the consensus is to
do this for GOP). The vast majority of the liveness reports that
we have gotten over the years in DBCP and pool are just pool
exhausted due to failure to return instances.
The workaround violating maxIdle will restore liveness, but is
likely the wrong solution here. Again, up to you guys to judge.
Note that when you specify maxIdle equal to 1
Crap. Meant 0 there.
you are telling the pool to destroy all returning instances. To
use the pool in this way is silly, IMO. With your patch, new
instances are still created for *every* borrow. The pool and all
of its machinery is doing nothing other than tracking the number
of active instances and making sure life cycle events are fired.
If you want the semantics of returning object with threads
waiting to be that the returning object is passivated, activated
and handed directly to a waiter without every hitting the idle
instance pool, that would require rewriting a fair bit of the
borrow / return code. It is an interesting idea and would solve
the POOL-240 as well.
One final comment. If you stick with something like what is in
the code now, you should make sure to passivate the new instance
before putting it into the pool. I just noticed that ensureIdle
does not do that, which I think is a bug in that method. So if
you want to proceed with this fix, I would recommend
1. Move the ensureIdle activations added in POOL-240 into
destroy itself.
2. Add passivation to ensureIdle
3. Implement corresponding workaround for GKOP
Sorry, all. Looking again and thinking a little more, I think the
best fix for this is actually to prevent the destroy on return in
the first place if there are take waiters. The maxIdle invariant
gets violated with the ensureIdle or direct create() fix anyway and
just allowing the returning instance to go into the pool avoids the
needless destroy / create sequence. A simple way to do this is just
to recode maxIdle before the destroy test in returnObject.
Instead of
final int maxIdleSave = getMaxIdle()
explicitly recode it:
// If maxIdle is set to 0 and there are take waiters, bump it to 1
// to preserve liveness.
final int maxIdleSave = idleObjects.hasTakeWaiters() ?
Math.max(1,getMaxIdle()) : getMaxIdle();
Phil
Phil
LieGrue,
strub
Am 23.11.2018 um 16:51 schrieb Phil Steitz<phil.ste...@gmail.com>:
On 11/23/18 2:57 AM, Mark Struberg wrote:
should read: This change (putting a new item back to the idle
pool) was needed to prevent a dead-lock....
*grabbing a fresh coffee* le
I am sorry I did not look carefully enough at this issue before
reviewing the change. After reviewing the DBCP ticket (OP
likely unrelated, IMO), I think that the right answer here is
WONT_FIX. That sounds harsh, but it seems to me that the
obvious solution here is for the user to set maxIdle to at
least 1. What the fix does is effectively that, without
changing the setting. If waiting threads die or time out while
the create is happening, there will be an idle instance in the
pool and for certain one is being put there on the way to
getting checked back out. See comments on POOL-327.
If the consensus is to insert this workaround to enable the
pool to retain liveness in the use case, it's probably best to
use ensureIdle(1, false) (see POOL-240). It could be that just
moving the call to ensureIdle inside destroy would be OK. But
as stated above, this breaks the maxIdle contract.
I see that your original report / use case here is from DBCP,
Mark. Was it prepared statements or connections that you were
trying to limit to 0 idle? Is there a reason that just using 1
would not work?
Phil
Am 23.11.2018 um 10:49 schrieb Mark Struberg<strub...@yahoo.de>:
This change (putting a new item back to the idle pool was
needed to prevent a dead-pool
---------------------------------------------------------------------
To unsubscribe, e-mail:dev-unsubscr...@commons.apache.org
For additional commands, e-mail:dev-h...@commons.apache.org
.
---------------------------------------------------------------------
To unsubscribe, e-mail:dev-unsubscr...@commons.apache.org
For additional commands, e-mail:dev-h...@commons.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail:dev-unsubscr...@commons.apache.org
For additional commands, e-mail:dev-h...@commons.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
