Am Dienstag, 11. August 2009 16:39:43 schrieb Johannes Meixner:
Hello,
On Aug 11 15:03 Adrian Schröter wrote (shortened):
Am Dienstag, 11. August 2009 12:45:57 schrieb Johannes Meixner:
On Aug 11 10:50 Adrian Schröter wrote (shortened):
Am Dienstag, 11. August 2009 10:34:58 schrieb Johannes Meixner:
One needs transaction semantics when working on packages: 1. Begin of transaction 2. Change it 3. End of transaction
This isn't there by intention. Basically no scm implements it because it blocks other people and slow downs development. Usually first submitter wins and followers need to adapt.
I can not imagine any system right now, which solves social problems, in the end people need always to talk to each other if they work on same code.
A weak synchronization point preferably with a kind of weak locking before someone starts to change it (so that all others who also work on it are at least informed) and a strong synchronization point preferably with a kind of three way merge when changes are committed.
I do not understand how a weak synchronization point before blocks other people and slow downs development.
The problem with the current build service is not the optimistic concurrency control policy at commit time, the problem is that nobody knows about others who may also work on a package until it is too late to start collaborating.
Remember Coolo's mass-changes of packages. All others notice what Coolo did only when it was already done. If there was a "begin of transaction" the others could get at least informed before.
I fail to see how a package locking would help here.
I never requested hard locks. A "Begin of transaction" does not require locks.
Transaction semantics is "all or nothing", i.e. either all changest between "Begin of transaction" and "End of transaction" are committed or nthing at all.
Transaction semantics works well with an optimistic concurrency control policy (i.e. first commit wins).
Transaction semantics with optimistic concurrency control would do:
1. Begin of transaction for package foo Remember the current write time stamps for all files of package foo
2. Read whatever files of package foo are needed and change a local copy of them.
3. End of transaction for package foo 3a) Do a hard lock for all files of package foo on the server 3b) Compare if a write time stamp for one of the files which was locally changed did already change on the server and if yes do not change anythjing on the server if no write the locally changed files onto the server and finally release the lock.
all this is there IMHO, just that usually md5sums are used instead of timestamps.
It would use by the script, so it would briefly lock the package and than submit it. Not much time in between.
Yes. The hard lock happens only while a transaction is at the end finally in it commit procedure.
No need for a lock, we just do it atomar already. And another commit fails if it is not based on the former revision.
Therefore we have the source links and devel projects, one get the changes (either automatically or by manually solving via repairlink).
But as you wrote - and as other threads here show - mutual information is not really within the scope of the build service but is left to the users of the build service to do it somehow on their own beside the build service via manual mails.
We can discuss to have something like a "I am currently working on this package" flag, but I strongly mind to block something.
Exactly what I ask for. I just want to be informed who else works on a package while I am working on it.
What I have in mind is a kind of "current workers list" on the server which contains who currently works on a package.
1. Begin of transaction 1a) Lock the current workers list (for read and write) 1b) Test if current workers list is empty 1b1) if empty, continue with step 1c 1b2) if not empty, show me the current workers list and ask me if I like to work on it additionally and wait (on the server) up to 1 minute for my response 1b2a) if response is "no" or if timeout, unlock the current workers list and cancel the transaction 1b2b) if response is "yes", continue with step 1c 1c) Add myself to the current workers list 1d) Send the current workers list to all current workers 1d) Unlock the current workers list 1f) Create local copy (i.e. "Read") 2. Modify local copy 3. End of transaction 3a) Lock 3b) Test if modifications do not conflict (i.e. "Validate") 3b1) if no conflict, write modifications (i.e. "Write") 3b2) if conflict, ignore modifications (loss of work) 3c) Unlock 3d) Remove myself from the current workers list 3e) Send the current workers list to me and all current workers
What is the timeframe between 1 and 3 ? 3. itself runs already atomar, so I do not see a reason for locking. Do you really thing this is needed, why can other developers with other scms survive without ? I mean, we even have source links and submit requests which allows parallel development, so what makes us so special that we need such a system ?
The two kind of "send the current workers list" messages should be processed by hermes (i.e. one can suppress them).
Perferably each worker could provide an optional notification which is also stored in the "current workers list" what the reason is why he works on a package like "fixing bnc#12345".
A worker is someone else have write access in your project or package, right ? IMHO a submit as often as possible approach would have already this effect, and you can already subscribe to source changes in your package in hermes.
...
In contrast it is crucial to have meningful "osc log" messages when I query whatever package on the server to find out which revisions are of particular interest regarding a particular issue.
what do you miss exactly ?
I do not mean the request messages but my "osc commit" messsages.
I would like to have the diff of the changes file as the minimum of meaningful "osc log" messages.
"osc help log" reads "Shows the commit log of a package" but currently it seems "osc log" does not show a log regarding what has changed inside a package (i.e. its changelog) but a log regarding what happened to a package in the obs i.e. whereto it was submited and if it was accepted and so on but according to my understanding this is not the commit log but a submit log.
Okay, the changelog handling is on our todo list since the begin of OBS. But it is a quite complex task. However, what is important to understand is that the commit log is target for the developer, which the changes file is target for the end-user. So the content of them might complete different. Nevertheless there should be of course a connection to avoid double work.
The use case which I have in mind is:
End-user jane runs out there whatever package from the obs. I get a bug report from her. I let her send me "rpm -q --changelog" and therein I find a suspicious entry:
* Wed Jan 14 2009 johndoe@somewhere - removed various obsolete stuff
Now I want to know exactly what johndoe did at that time.
My first problem is to find out the osc revision which matches to this RPM changelog entry. I don't know a direct way to get it.
You even have more problems, you don't know from which project (or even package) this is build from and you don't know which code version it is exactly (there might be multiple versions with this last changelog entry). Last but not least, you don't know if it is really from this build service instance. We have introduced therefore the disturl tag sicne SLE 11. So please ask your users for the output of it: # rpm -q --qf '%{DISTURL}\n' osc obs://build.opensuse.org/openSUSE:Tools/openSUSE_Factory/8f3b8001f49bc9e991b80cf0d7cfb5c4- osc you can check out the code now via # osc co openSUSE:Tools osc -r 8f3b8001f49bc9e991b80cf0d7cfb5c4 or call rdiff to see the changes since then. bye adrian -- Adrian Schroeter SUSE Linux Products GmbH email: adrian@suse.de -- To unsubscribe, e-mail: opensuse-packaging+unsubscribe@opensuse.org For additional commands, e-mail: opensuse-packaging+help@opensuse.org