Hello Elena, You raise good points. I have just rewritten the save_state and load_state functions. Now they work with a MySQL database and a table that looks like this: create table kokiri_data ( dict varchar(20), labels varchar(200), value varchar(100), primary key (dict,labels)); Since I wanted to store many dicts into the database, I decided to try this format. The 'dict' field includes the dictionary that the data belongs to ('upd_count','pred_count' or 'test_info'). The 'labels' field includes the space-separated list of labels in the dictionary (for a more detailed explanation, check the README and the code). The value contains the value of the datum (count of runs, relevance, etc.) Since the labels are space-separated, this assumes we are not using the mixed mode. If we use mixed mode, we may change the separator (, or & or % or $ are good alternatives). Let me know what you think about this strategy to store into the database. I felt it was the most simple one, while still allowing to do some querying on the database (like loading only one metric or one 'unit' (platform/branch/mix), etc). It may also allow to store many configurations if necessary. Regards Pablo On Sat, Aug 9, 2014 at 8:26 AM, Elena Stepanova <elenst@montyprogram.com> wrote:
Hi Pablo,
Thanks for the update. Couple of comments inline.
On 08.08.2014 18:17, Pablo Estrada wrote:
Hello Elena, I just pushed a transaction, with the following changes:
1. Added an internal counter to the kokiri class, and a function to expose it. This function can show how many update result runs and prediction runs have been run in total, or per unit (an unit being a platform, a branch or a mix of both). Using this counter, one can decide to add logic for extra learning rounds for new platforms (I added it to the wrapper class as an example).
2. Added functions to load and store status into temporary storage. They are very simple - they only serialize to a JSON file, but they can be easily modified to fit the requirements of the implementation. I can add this in the README. If you'd like for me to add the capacity to connect to a database and store the data in a table, I can do that too (I think it
Yes, I think we'll have to have it stored in the database. Chances are, the scripts will run on buildbot slaves rather than on the master, so storing data in a file just won't do any good.
would be easiest to store the dicts as json data in text fields). Let me
know if you'd prefer that.
I don't like the idea of storing the entire dicts as json. It doesn't seem to be justified by... well... anything, except for saving a tiny bit of time on writing queries. But that's a one-time effort, while this way we won't be able to [easily] join the statistical data with, lets say, existing buildbot tables; and it generally won't be efficient and easy to read.
Besides, keep in mind that for real use, if, lets say, we are running in 'platform' mode, for each call we don't need the whole dict, we only need the part of dict which relates to this platform, and possibly the standard one. So, there is really no point loading other 20 platforms' data, which you will almost inevitably do if you store it in a single json.
The real (not json-ed) data structure seems quite suitable for SQL, so it makes sense to store it as such.
If you think it will take you long to do that, it's not critical: just create an example interface for connecting to a database and running *some* queries to store/read the data, and we'll tune it later.
Regards, Elena
By the way, these functions allow the two parts of the algorithm to be called separately, e.g.:
Predicting phase (can be done depending of counts of training rounds for platform, etc..) 1. Create kokiri instance 2. Load status (call load_status) 3. Input test list, get smaller output 4. Eliminate instance from memory (no need to save state since nothing changes until results are updated)
Training phase: 1. Create kokiri instance 2. Load status (call load_status) 3. Feed new information 4. Save status (call save_status) 5. Eliminate instance from memory
I added tests that check the new features to the wrapper. Both features seem to be working okay. Of course, the more prediction rounds for new platforms, the platform mode improves a bit, but not too dramatically, for what I've seen. I'll test it a bit more.
I will also add these features to the file_change_correlations branch, and document everything in the README file.
Regards Pablo
On Wed, Aug 6, 2014 at 8:04 PM, Elena Stepanova <elenst@montyprogram.com> wrote:
(sorry, forgot the list in my reply, resending)
Hi Pablo,
On 03.08.2014 17:51, Pablo Estrada wrote:
Hi Elena,
One thing that I want to see there is fully developed platform mode. I
see
that mode option is still there, so it should not be difficult. I
actually
did it myself while experimenting, but since I only made hasty and crude
changes, I don't expect them to be useful.
I'm not sure what code you are referring to. Can you be more specific on what seems to be missing? I might have missed something when migrating
from
the previous architecture...
I was mainly referring to the learning stage. Currently, the learning stage is "global". You go through X test runs, collect data, distribute it between platform-specific queues, and from X+1 test run you start predicting based on whatever platform-specific data you have at the moment.
But this is bound to cause rather sporadic quality of prediction, because it could happen that out of 3000 learning runs, 1000 belongs to platform A, while platform B only had 100, and platform C was introduced later, after your learning cycle. So, for platform B the statistical data will be very limited, and for platform C there will be none -- you will simply start randomizing tests from the very beginning (or using data from other platforms as you suggest below, which is still not quite the same as pure platform-specific approach).
It seems more reasonable, if the platform-specific mode is used, to do learning per platform too. It is not just about current investigation activity, but about the real-life implementation too.
Lets suppose tomorrow we start collecting the data and calculating the metrics. Some platforms will run more often than others, so lets say in 2 weeks you will have X test runs on these platforms so you can start predicting for them; while other platforms will run less frequently, and it will take 1 month to collect the same amount of data. And 2 months later there will be Ubuntu Utopic Unicorn which will have no statistical data at all, and it will be cruel to jump into predicting there right away, without any statistical data at all.
It sounds more complicated than it is, in fact pretty much all you need to add to your algorithm is making 'count' in your run_simulation a dict rather than a constant.
So, I imagine that when you store your metrics after a test run, you will also store a number of test runs per platform, and only start predicting for this particular platform when the count for it reaches the configured number.
Of the code that's definitely not there, there are a couple things that could be added: 1. When we calculate the relevance of a test on a given platform, we
might
want to set the relevance to 0, or we might want to derive a default relevance from other platforms (An average, the 'standard', etc...). Currently, it's just set to 0.
I think you could combine this idea with what was described above. While it makes sense to run *some* full learning cycles on a new platform, it does not have to be thousands, especially since some non-LTS platforms come and go awfully fast. So, we run these no-too-many cycles, get clean platform-specific data, and if necessary enrich it with the other platforms' data.
2. We might also, just in case, want to keep the 'standard' queue for
when
we don't have the data for this platform (related to the previous point).
If we do what's described above, we should always have data for the platform. But if you mean calculating and storing the standard metrics, then yes -- since we are going to store the values rather than re-calculate them every time, there is no reason to be greedy about it. It might even make sense to calculate both metrics that you developed, too. Who knows maybe one day we'll find out that the other one gives us better results.
It doesn't matter in which order they fail/finish; the problem is, when
builder2 starts, it doesn't have information about builder1 results, and builder3 doesn't know anything about the first two. So, the metric for
test
X could not be increased yet.
But in your current calculation, it is. So, naturally, if we happen to catch the failure on builder1, the metric raises dramatically, and the failure will be definitely caught on builders 2 and 3.
It is especially important now, when you use incoming lists, and the running sets might be not identical for builders 1-3 even in standard
mode.
Right, I see your point. Although if test_run 1 would catch the error, test_run 2, although it would be using the same data. might not catch the same errors if the running set makes it such that they are pushed out due to lower relevance. The effect might not be too big, but it definitely
has
potential to affect the results.
Over-pessimistic part:
It is similar to the previous one, but look at the same problem from a different angle. Suppose the push broke test X, and the test started failing on all builders (platforms). So, you have 20 failures, one per
test
run, for the same push. Now, suppose you caught it on one platform but
not
on others. Your statistics will still show 19 failures missed vs 1
failure
caught, and recall will be dreadful (~0.05). But in fact, the goal is
achieved: the failure has been caught for this push. It doesn't really matter whether you catch it 1 time or 20 times. So, recall here should
be 1.
It should mainly affect per-platform approach, but probably the standard one can also suffer if running sets are not identical for all builders.
Right. It seems that solving these two issues is non-trivial (the
test_run
table does not contain duration of the test_run, or anything). But we can keep in mind these issues.
Right. At this point it doesn't even make sense to solve hem -- in real-life application, the first one will be gone naturally, just because there will be no data from unfinished test runs.
The second one only affects recall calculation, in other words -- evaluation of the algorithm. It is interesting from theoretical point of view, but not critical for real-life application.
I fixed up the repositories with updated versions of the queries, as
well as instructions in the README on how to generate them.
Now I am looking a bit at the buildbot code, just to try to suggest some design ideas for adding the statistician and the pythia into the MTR related classes.
As you know, we have the soft pencil-down in a few days, and the hard one a week later. At this point, there isn't much reason to keep frantically improving the algorithm (which is never perfect), so you are right not planning on it.
In the remaining time I suggest to
- address the points above; - make sure that everything that should be configurable is configurable (algorithm, mode, learning set, db connection details); - create structures to store the metrics and reading to/writing from the database; - make sure the predicting and the calculating part can be called separately; - update documentation, clean up logging and code in general.
As long as we have these two parts easily callable, we will find a place in buildbot/MTR to put them to, so don't waste too much time on it.
Regards, Elena
Regards Pablo