Re: How to handle a migration that last too long for being deployed?

Hi,

On 04/14/2018 02:38 AM, Mike Dewhirst wrote:

Does it actually stop users reading? If the entire migration happens in a single transaction, the database (Postgres anyway) should remain accessible until the moment it is committed.

Maybe you could announce a maintenance operation which will only interrupt certain actions for a few minutes?

I am not sure if I understand well how that works. If the migration is atomic, is is true that:
- the users can read normally between the beginning of the migration and the beginning of the transaction commit, and they will get the old data
- the users trying to read during the transaction commit will have to wait for it to finish, and they will get the migrated data
- the users who tries to write at anytime between the beginning of the migration and the end of the transaction commit will have to wait for it to finish, and they might overwrite the migrated data

If it works like that, I think that the solution I was thinking about is good enough for my needs:

- process the queryset by chunks. Note that if you want to use prefetch_related, you can't use a queryset iterator. So I get the successive chunks by filtering with PK ranges ... I have benchmarked a bit and a good value for the chunk size seems to be 500, it's not slower then any other value and keeps the memory usage down. If I had a lot of RAM, I could also raise this value to 5000 or 15000 without really slowing the process.

- I have actually two separate atomic migrations (in case something goes wrong, it is better then one non atomic migration with two atomic operations): one to process the objects that are not modifiable by the users (because they are in a certain status etc.) and a second to process the objects that could be modified by the users. The second migration concerns only 4-5% of the total objects, so it should be much faster. As I use the PKs to fetch the objects, I have to fix the desired chunk size of 500 in order to get some chunks of (approx.) the same size with 500 * total_objects_count / filtered_queryset_count.

There remains the problem of users that would try to write during the second migration: their changes will be written indeed to the old model, but not taken in account by the new models (remember I want to split one model in two smaller ones). So maybe I should append here to the second migration all the operations that are responsible for deleting the old model? This way, people trying to write will get an error - which is the best we can do here. Am I right?

Thanks for your help!

Cheers,
Adrien

Adrien Cossa <cossa@init.at> wrote:

Hi everybody!

I would like to know what options exist when you have a huge migration that will obviously not run on your productive server.

I have spitted a model in two smaller ones and wrote then a migration to populate these new models. The number of original objects is around 250,000 and I have also a few references to update. In the end, the migration lasted more than 30 mn on my machine (16 GB RAM and it was swapping a lot) and it failed on another machine because the RAM was out (the process was using then about 13 GB). On the productive server we have even less RAM so to run the migration as it is is really out of question.

I have tried to use all the Django mechanisms that I know to optimize the queries: select_related, prefetch_related, bulk_create, QuerySet.update... Now, the migration I am talking about use bulk_create(batch_size=None) and process the whole queryset at once. Before that, as the migration was not so long lasting because I had 2 references less to update,  I tried other values for batch_size and also I processed the queryset as pages of a few hundreds or thousands objects. The results were not better then batch_size=None and "all at once", that's why I finally used "basic settings" (and the migration was lasting about 5 minutes). I will have to reintroduce some tweaks because the extra updates of the two relations I mentioned is making here a big difference.

I am wondering if someone already found him/herself in a similar situation, and with what solution you finally came to.

If the migration lasts very long, it's not a problem by itself but I don't want to lock the database for 15 mn. The fact is that I don't know what is happening during the migration process, what is locked by what? I will split the migration in "pages" to use less RAM anyway, but I was also thinking of migrating in two different steps *or* files, in order to process separately the objects that are not editable (basically most of them, that we keep for history but they are read-only) and the others (which should be much faster and thus people working will not be blocked for long). Does it make sense? Some other ideas?

Thanks a lot!

Adrien