You are right about the getData, and it was probably a wrong choice.
Both the getData and getChildren have an extra ACL check in the
FinalRequestProcessor. My question is pretty much this: shouldn't the
other request types have such a check as well? What if client B has
issued a setData()? setData is only ACL-checked in PrepRequestProcessor,
so it may be the case that it succeeds, while client A's the setAcl()
has been ordered to be executed first.
It seems that an ACL check at the FinalRequestProcessor for most request
types would solve this issue. As it is there only for getData and
getChildren, I wonder why that is and if scenarios as the one above are
By the way, does Zookeeper require clients to wait for the reply for
setAcl()? If not, then the "single client, single server" model would
face the problem even more obviously.
Thank you for the replies,
Benjamin Reed wrote:
in your scenario the getData will not succeed. the acl will be checked in the
FinalRequestProcessor after the setAcl has passed through the final request
processor. execution only happens in the final request processor and always
happens in order.
the only way a getData will succeed after a setAcl succeeds if it is done on a
lagging follower, but in that case it will see a lagging view of the system.
consider the following:
/foo has an open ACL
client A: setAcl("/foo", restricted ACL); setData("/foo", new data);
client B: getData("/foo");
no matter what happens, what the delays are, if B is faster or slower than A, client B
will not see "new data". it will either pass the ACL check and see the old data
or it will fail with a permission exception.
From: Manos Kapritsos [ma...@cs.utexas.edu]
Sent: Tuesday, February 10, 2009 4:34 PM
Subject: Re: setACL semantics
More or less, yes.
If the requests are from the same client, then you say that there will
not be a problem? I guess that is true if you always wait for the
response of the first request in order to execute the second. I am not
sure if that is a requirement for all Zookeeper client implementations.
As for two different clients (which was the case I was thinking about),
this seems to be a problem. I will agree that (if clients only have one
outstanding request) the two requests are concurrent and either order of
execution is considered to be acceptable, but it could be that two
different replicas receive the two requests in the same order, but
effectively execute them in a different order. In any case, it feels
wrong (at least to me) that a getData would succeed when a setACL that
prevents it has already been accepted to be processed.
Mahadev Konar wrote:
If the setAcl and getdata are from the same client then they are all
handled in order. So you would get an unauthroized exception when you do a
If two diff clients do setacl and getdata it might be that the getdata in
your case will succeed before the setacl returns on the first client.
Is that what you meant?
On 2/10/09 2:15 PM, "Manos Kapritsos" <ma...@cs.utexas.edu> wrote:
I have a question about the way setACL functions. It seems that the
PreRequestProcessor handles all kinds of requests the same, checks the
validity of the corresponding ACL, and enqueues them to Sync and Final
processors. Maybe I am missing something here, but this behaviour seems
weird. What if a setACL request comes, setting the ACL of a path (e.g. /
) to an IP (e.g. 18.104.22.168) , instead of its old value (e.g. World).
This request will pass the ACL check, and will be enqueued to be
processed by the next processors. Assume that the next request is a
getData("/") from an IP other than 22.214.171.124. If this request is
processed by the PreRequestProcessor before the setACL request is
processed by the FinalRequestProcessor, then it will pass the ACL check
(which it should not, since it came after the setACL request). It seems
that there is a race condition here that should not exist.
Let me know if this is actually the case or I am missing something. I am
using version 3.0.1 of the code.