Huge News!Announcing our $40M Series B led by Abstract Ventures.Learn More
Socket
Sign inDemoInstall
Socket

wyseman

Package Overview
Dependencies
Maintainers
1
Versions
36
Alerts
File Explorer

Advanced tools

Socket logo

Install Socket

Detect and block malicious and high-risk dependencies

Install

wyseman

PostgreSQL Schema Manager with Javascript, Ruby, TCL API

  • 1.1.13
  • latest
  • npm
  • Socket score

Version published
Weekly downloads
42
increased by740%
Maintainers
1
Weekly downloads
 
Created
Source

Wyseman: WyattERP Schema Manager

Wyseman is a database schema manager, part of the WyattERP application framework.

It is used for authoring and managing database schemas in PostgreSQL. It also provides run-time access to a dictionary for access to meta-data about database objects.

Run-time support includes:

  • Javascript
  • Ruby
  • Legacy TCL

This is how it works. You author your SQL objects inside TCL data containers like this:

    table myschema.mytable {dependency1 dependency2 ...} {
      columnA	typeA 	primary key,
      columnB	typeB,		-- Regular SQL stuff in here
      columnC	typeC		-- Just as in 'table create'
    }

This tells wyseman what your object is, how to create it, and unless it is self-evident, how to drop it. Most importantly, it tells if there are other objects that will have to be recreated as well if we want to rebuilt this one.

There are two other very similar object decriptions you can make too:

  1. A tabtext contains language data about the table, columns, and possible values for your data. You can define this for any number of different languages. For example:
  tabtext myschema.mytable {What my table is for} {
    columnA	{Title for A} {More detailed info for column A}
    columnB	{Title for B} {More detailed info for column B}
    columnC	{Title for C} {More detailed info for column C} {
      value1	  {Title for 1} {What this possible C value means}
      value2	  {Title for 2} {What this possible C value means}
      value3	  {Title for 3} {What this possible C value means}
    }
  } -language en
  1. A tabdef contains information about how your tables and views should normally render and otherwise, be handled, in the user's view.

Once you have your schema defined, just tell wyseman to build it from your files. It will store all your objects into a special schema in the database. And on your command, it will build the target schema you have just designed.

Then, as you make changes to your object definitions, wyseman can modify and upgrade the running database, without disrupting any data it may contain, to keep it current with your design.

Now go create your frontend using Wylib and your application will automatically have access to all the information about your database tables, columns and so forth.

There is also an associated package called Wyselib that contains a bunch of basic schema components for handling common things like people, accounts, products, etc.

These packages work together to make it relatively painless to make and more importantly, maintain a production database application.


Background History

When I first started using SQL relational databases, there was one issue I continually struggled with: How is it best to maintain the source definition of my database schema.

I started out editing my SQL in neatly commented text files. During the testing phase, I would periodically drop the database and re-create it from my text files until I finally had a working system.

This all worked out very well until I began to find the weaknesses in my design and wanted to fix them. Now the problems was, I had live data in my tables so I couldn't just drop the database and start over. I had to alter the database without disturbing it too much so it could stay in operation for my users.

Like most people, I discovered the "alter" commands of SQL and began carefully adjusting tables, adding columns, altering views, changing functions etc. The problem was, my original SQL files quickly began to become obsolete. They described the schema I had originally crafted, but not the one it was becoming.

For a while, I tried to keep them current. Each time I would alter the database, I would edit my SQL files to reflect the changes I had made. But these changes never actually got tested so I was never quite sure if I had gotten them right.

In one case, I found that in dropping and re-creating certain tables, I had inadvertently dropped certain key referential integrity triggers. This resulted in a large amount of bad data creeping into my database and long hours of work to fix it.

In another case, I wanted to re-create my schema for another enterprise. When I tried to run my SQL files, I found indeed they were full of errors. Even after fixing the errors, I was not quite sure they reflected my working production system.

I tried using pg_dump to dump out the schema from my production system and re-create it on the new system, but I was frustrated that the SQL was really not the original code I had written, but a bit more cryptic form. I was able to re-create the database exactly, but it proved difficult to modify it for the needs of the new enterprise except through a set of new alter commands.

It finally occurred to me that the create/alter model of SQL assumes you are comfortable with the notion that the source for your model is contained within the working instance of the database yourself. Any text files you create are merely temporary representations of the modifications you apply. The model consists of the net result of all the changes that have been applied over time. If you want to replicate or alter the model, you are limited to the alter/dump commands available and you'd better get it perfect as you apply your changes to the production database.

As I used SQL more, I found myself using similar structures over and over again. For example, when adding insert or update rules to views, I used similar syntax over and over with a few key parameter changes. Sometimes, I also found other similar table or function structures that could have been re-useable if only I had some kind of macro or scripting capability on top of SQL.

Eventually, I set out to create a tool for managing my schema definitions that would help me solve these and a few other problems.

I determined it would be best to have my source model contained in text files rather than in the running instance of the database. This would give me more control over my authoring content. If I wanted to alter the model or create new instances, I could do it at any time regardless of the accessibility of my production instance(s).

By using text files, this also opened up the opportunity to use pre-processing (macros, functions, etc.) to make my schema definitions more concise and maintainable.

The original attempt consisted of a handful of shell scripts and a syntactical convention for placing bits of SQL inside my schema files in predictable ways so the shell scripts could execute the right bits of SQL at the right times.

Certain sections of the files contained the code for dropping certain objects. Other sections had the SQL to create those same objects. Objects were also categorized by type which led to the very handy feature of being able to drop and recreate all objects of a certain kind (like re-indexing the whole database from a single command.)

I used m4 to pre-process my files, so I could begin to wrap bits of code in macros. This allowed me to create more SQL from a smaller set of source code. The result was, certain kinds of changes only needed to be made in one place and the changes would propagate throughout the database.

I also had a method by which certain source files were dependent upon others. This assured that the database objects were build (and/or dropped) in the right order.

All this worked so well, I decided to take another crack at it and see if I could make it even better. I decided to make it a part of my open-source ERP toolkit Wyatt-ERP and added a data-dictionary capability for storing more information about tables, columns and column values.
The result was Wyseman.

I very much liked the scripting capabilities of Tcl. M4 had a number of problems that worked out much better in Tcl. The Tcl list constructs also proved to be a great way to package the SQL bits and organize them into the proper dependencies.

With this new implementation of Wyseman, it became possible to create an entire database quickly and easily. With an appropriate Makefile, this could be done with a single command. Because everything is scripted in Tcl, you could even parameterize your schema (i.e. you could create schemas that are customized for different instances, but which are described by a single set of schema files.)

Once you got your instance running, it was much easier to operate on the "living patient." You could select any branch of the "tree" of dependencies (like a function, for example), and with one command, you could drop the old version of the object with all its dependencies, and then rebuild it again fresh from your schema definition files.

This gives the assurance that your running instance is truly the product of your schema definition files. And you don't have to worry about whether all the dependent objects got correctly rebuilt or not.

If any of the objects you need to rebuild are tables, the data would be dumped to a file before you did the drop, and re-inserted after the table is re-created. As long as the same columns exist in the tables before and after, this all works.

So if you need to change the columns of a table, you would have to write a shell script that did the following:

  1. Correct any known anomalies in the data
  2. You can safely drop any data integrity triggers if needed
  3. Issue alter commands to add/drop table columns
  4. Synthesize appropriate data for any newly added columns
  5. Call wyseman to drop/restore the applicable database sections

Before you run your script on the production instance, you would use pg_dump to create a sandbox copy of the instance and test your upgrade script on that. When everything looks good, you can shut down access to the production instance and run the upgrade for real.

This all worked pretty well and served all the needs of the business it was created for.

Fast-forward 10 years or so and I'm working on a new project that needs these same kind of capabilities. However it was clear, Wyseman needed a bit of sprucing up to meet the needs of this project. So the current iteration includes the following upgrades:

  • Reimplemented the command line program in Ruby, rather than TCL
  • Improved the amount of data contained in the data dictionary
  • Created the notion of schema versions to track multiple releases
  • The schema files are still authored in text files, but then cached in the database itself, where the drop/re-create can all happen in a single transaction now.
  • Implemented a Ruby run-time library to access the meta-data
  • Implemented a Javascript run-time library to access the meta-data

Keywords

FAQs

Package last updated on 19 Sep 2023

Did you know?

Socket

Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.

Install

Related posts

SocketSocket SOC 2 Logo

Product

  • Package Alerts
  • Integrations
  • Docs
  • Pricing
  • FAQ
  • Roadmap
  • Changelog

Packages

npm

Stay in touch

Get open source security insights delivered straight into your inbox.


  • Terms
  • Privacy
  • Security

Made with ⚡️ by Socket Inc