Re: Managing database change: improvements needed

I thought I sent my response but I guess something went wrong. Sending again.

Remarks in-line

On Wed, Aug 26, 2015, 14:23 Erik Huelsmann <..hidden..> wrote:

Hi all,

There are a few issues that I have with the way we currently handle database schema changes.

Current situation

============

We have a main schema file for the table definitions and a number of modules with stored procedures grouped by "subject".

Next to that, there's a "Fixes" file, which contains all incremental updates to the schema since some version.

When I want to change the schema as part of our development, I need to change the schema definition files *and* I need to add the schema changes to the Fixes.sql file.

I agree this is less than ideal. I think we need to think of a comprehensive tooling approach and there are a few reasons we haven't done this (mainly because I think getting the tooling right is a lot of effort and there has in the past been other urgent things that needed more effort at the beginning). However I think this is growing in urgency as we continue development.

Problems with the current approach

==========================

The current approach has (IMO) several problems:

* When loading the "Fixes.sql" file, even in a production environment, produces a truckload of ERRORs, because the schema may contain some or all of the fixes in the file, leading to aborted transactions

* Making adjustments in multiple places can lead to incompleteness on either side

* Fixes.sql is loaded twice; before and after loading the stored procedures, making it unsuitable for some kinds of fixes

* In order to change types or parameter lists of stored procedures, they need to be dropped before being (re)created, resulting in DROP TYPE and DROP FUNCTION all over the place.

Especially the high number of ERRORs in the log file when creating a new database is disconcerting to new users. Myself, I'm running into the DROP TYPE and DROP FUNCTION more and more often as my refactorings for 1.5 and 1.4-mc reach deeper into the system.

Additional requirements

=================

* Next to solving the problems above, some users have expressed the desire to extend LedgerSMB locally. However, the way we currently upgrade the system removes all customizations (at least from 1.2->1.3->1.4; maybe not 1.4->1.5). If our changed approach could help solve this requirement, all the better.

* The solution chosen preferably works with a strategy of branching and merging as we do in the development cycle.

Agreed on all the above.

Possible solutions

=============

One thing that I think we should do is separate development from production when loading the database. That is to say: when creating a new database, that is a different step than upgrading a production database which is in turn again very different from working on a development database. Our processes should work to provide the best for each.

I am not sure I understand what you mean by this. Do you mean that we need an extra build step? I would not be opposed to that. But if that isn't what you mean, could you be more specific?

Proposed solution to the duplicate change problem

=====================================

In the past we have been talking about adding a "build" step to LedgerSMB. This step could be used to produce a single schema file for quickly setting up a new database (company). This schema file would be a build artifact and no longer be version controlled itself. It'd be the result of a full schema creation step, including running the Fixes.sql file.

Additionally, this step could be used to deliver an up-to-date doc/database/ directory with current database documentation.

I like that idea.

While this step may feel inhibiting for development, one thing I'm thinking is that we may not need to require this step to be executed on a development system, except for when testing the production deployment.

And testing the build system.

Proposed solution to the slew of errors from Fixes.sql

=======================================

There are actually a number of solutions here, as I see it, all of them revolving around the idea that every schema change should be applied once. Basically, I see 2 categories of solutions:

1. Do it ourselves

2. Re-use the work of others

The benefit of (1) is that we get full control and no further dependencies for development or production. However, the downside is that we get to do all the thinking and problem solving as well.

In both categories I see similar solutions available:

a. Numbered changes

b. Named changes

and in (1) I see a solution that's not available in category (2):

c. Use PostgreSQL EXTENSIONs

As for category (2), I haven't looked too far around yet, but I did find sqitch (http://sqitch.org/); Sqitch offers a command line tool for the development workflow. Additionally, it provides facilities for deploying only the necessary changes with releases *and* it provides an API which we can use to upgrade the database from one (patch) release to another.

PG Extensions are definitely in the "reuse work of others" category but the system solves a different problem than sqitch. Indeed the two together could be a very powerful combination.

PG Extensions would allow us to ship db components (whether table sets, etc) in a versioned way. There is likely to be a learning curve here and it adds a major build dependency for people installing from our source packages. But it provides some really nice features including versioning and dependency tracking/enforcement.

What PG Extensions would allow us to do would be to effectively ship versioned pieces of our db schema and have an orderly way of upgrading them. For things like stored procedure sections, we could always drop the extension and load a new version. But it would make it very clean. And we could even break up parts of the db into modules that could be independently versioned.

Getting this tooling to work right is not a minor project however. It definitely cannot be done in a minor release because of the additional dependencies, and I think we'd need a policy against adding extensions to core in major versions. However it would also provide a nice platform for third party integration and reuse as well.

Each of our modules could be an extension and separately versioned.

For branching and merging, the extensions would have a number specific to development head and could be re-numbered during the build process. For modules all we really need to do is drop extension/create extension if the version number has changed.

We'd probably want to reserve a number series for "next in version series." I am thinking that using 999 would be a good one, so 1.4.999 would be "whatever next comes in 1.4" and this would allow us to install and run out of the development tree.

As for sqitch, I have no experience with it yet. It's supposed to work well with branching and merging. One thing it *does* do is integrate with version control tools and it means to integrate with the project's repository. Ideally when merging and branching, no additional processing is required to integrate the changes from branches. However, I'm not exactly sure that's what happens (given https://groups.google.com/forum/#!searchin/sqitch-users/merge/sqitch-users/GXqgt7nJ_1k/Vvg-r1HOEqMJ) but I think the referenced link is about reordering changes which already have been partially deployed. What I like about Sqitch is that it integrates with the VC system, but is VC system agnostic, some of us can use it with Git while others can keep using Hg-git as they currently do.

Ok so if we ship our storage schema as an extension (or group of extensions) we still need to have ordered versioning going on.

I am thinking that sqitch may allow us to do the grunt work of the schema change scripts as a part of the build phase (we have to strip out begin/commit statements since create/alter extension adds those), but it seems like it would be very helpful in managing that side.

What I *am* sure about is that (1) will be completely change-oriented. *Maybe* we can use sqitch with a the current module-based source code organization.

If we go with extensions we don't need to. We just bump the version number on each change of a module and keep a catalog of what is required (and tests to verify of course). We can then drop/create extensions as needed for the procedural modules, and we can generate migration scripts for the for the storage extension(s). The big difference is that instead of one db version we would have a bunch of components that were versioned and likely a requirement that these match on login.

If we're going to go with Sqitch, I don't think it's a good idea to switch just then and there, but test it to build up some experience and then choose a well defined point in time to start using it.

Agreed. Maybe start on plugins?

So, there are 2 proposals here. What do you say?

I don't see how we can get away from doing some things ourselves, so I guess the question is what and where to reuse other work.

So my preference would be sqitch and pg extensions. In my mind this would require a two-step build process for dbs:

1. Build the schema extensions using sqitch

2. Install all pg extensions using make.

The first is a developer-only task (sort of like running YACC is when building PostgreSQL -- not even required when installing from source packages) so there isn't a user-visible dependency there.

The second though does impose some user dependencies in that pg_config and other pg development tools must be installed before the build can be done. Of course for those installing debs, these would be done when building debs rather than when installing them.

So that is my preference and why.

--

Bye,

Erik.

http://efficito.com -- Hosted accounting and ERP.

Robust and Flexible. No vendor lock-in.

------------------------------------------------------------------------------
_______________________________________________
Ledger-smb-devel mailing list
..hidden..
https://lists.sourceforge.net/lists/listinfo/ledger-smb-devel