Six Ways to Build Reliable Applications

I am perpetually frustrated by the big-IT solutions that come out of the software industry under the premise that "it has to be big to do all it does" and the belief that "if it's not big, it's not a world class solution". Think about it:  a few things will come to mind. Like some database vendors, business process software solutions, and, yes, CM solutions. In the latter area, the industry has seen a few small-IT tools emerge which meet the requirements better than the big-IT tools, at least in some areas for some tools, in a broader sense for at least one. Better small-IT solutions will emerge over time across the entire software spectrum.

When I buy a software solution, I don't want an internal help desk, a need for significant consulting, a huge training bill, a large platform resource requirement, a solution administration team. I want ease of use, easy customization, small footprint, zero administration, reliability and loads of features that meet my requirements, out of the box if possible. This is certainly true of CM tools as well.

Architecture First
First and foremost, in building reliable applications, these things have to be the minds of the architects. The software industry aims too low in general. Let's glue things together. Let's copy and paste and change it. We'll optimize it later, or we can use bigger boxes to run it. Let's get it running first, then we can look at the features we need.

If you're in this type of environment, I don't envy you. What's missing is the building of the underlying architecture. Some might say architecture doesn't just happen. But more properly, good architecture doesn't just happen. If you don't build it well, you'll have a poor architecture.

The philosophy that says "just make it work" is very pervasive across the software industry, largely because it's easy to do and there are so many writing software. This is not to mention the schedule pressures to get to market. If you want reliable applications, you need to have good architects and good architecture. You have to replace the "pick one of the many ways to do it" philosophy with the "what's the proper way to do this". And to answer that question, you need to understand architecture from the ground up and develop the appropriate rules to guide you. There may be a few in your midst that do - maybe it takes them a couple of days to rationalize their "gut feel" about something. But it's usually worth it.

Take a look at various large software systems - how rapidly do they evolve? The ones that evolve slowly, or perhaps not at all may have maxed out on their architecture (though this is not always the reason they don't evolve). From the bottom up, the software must be well architected. If you're cutting and pasting code all over the place, it becomes much more than a challenge to universally improve the functionality of that portion of code.

" alt="" src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif" data-mce-src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif"

Re-Use
After designing the overall architecture, build fully re-usable libraries/APIs from the bottom up. Design the interfaces and review them with system architects for maximum re-useability and for appropriate usability.

For example, file systems have always had re-useable APIs, but not useable ones. "Set this flag on and that one off" and call it with this action and the routine can do about anything you want. I want simple to use routines, and so I always start a major project by introducing a file system API. It has meaningful routines such as:

·         FindFile

·         CreateFile

·         OpenFile

·         Seek

·         ReadBytes

·         WriteBytes

·         OpenTextFile

·         AppendToTextFile

·         TextRead

·         TextWrite

·         CloseFile

Then a bunch of directory functions and a few other file functions in similar vein.

I don't have to set flags to say that working with binary data or ascii data. I don't have to set modes saying "create if and otherwise open" or vice versa. I don't have to do different read operations for text and binary data using the same functions. That's all taken care of and implicit in the calls I use. Even more to the point, the functions are easy for anyone to pick up, and make for clearer code reading. OK, there are a few good file system APIs out there, but it's just that the universal ones seem to prefer assembly language level usability. And I don't mind having the ability to do raw IO on devices. But please don't impose that level on everyone!

" alt="" src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif" data-mce-src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif"

Having our own API also allows us to easily add functionality to our file system, such
as using "> []" to redirect output to a window. And, it provides a nice re-usable module which is portable across platforms (by design).

We know when we start a major project that we'll need to: access the file system, have a database, deal with storage allocation issues, etc. So we start by ensuring we have the "best" APIs for these functions - APIs that are easy to use and which are highly functional. And because they're re-usable, we benefit more than once. Not only do we have these advanced, portable modules ready for any project, but the reliability of them has already been tested out on other projects.

We don't stop with a few module APIs, we go much further. For example, at Neuma we have modules for command line interpreters, name directories, database query languages, macro definition and substitution, networking, GUI generation, and many more. Most of these were largely developed in the 1980's and 1990's, but they still evolve to support, for example, new GUI architecture advances. After a couple of decades, they have become very stable and very reliable.

In effect, generating these libraries of re-usable APIs allows us to build new applications very rapidly. We're basically using a very high level language that is even beyond that of  Perl or Visual Basic. That's our goal, to keep making it easier to build new applications, even to the point of exporting this capability to the run-time environment.

A Glimpse Inside Neuma
So how does Neuma do design? Or better yet, how did Neuma build a reliable CM product? What did we do right and what did we do wrong?

First, we decided to address areas where we were very strong. These were primarily Database and Configuration Management, where I as founder had had over a dozen years of experience in helping to architect a couple of major Telecom products which were highly successful. Even before we decided to focus on CM, Neuma began building a Next Generation Database (NGDB) because we saw a need for it in the market-place and we were good at doing it. We especially saw a need for a small-IT database.

We focused, as we did in the telecom industry where 4 hours down time in 40 years is acceptable, on reliability. Reliability has to be planned into a system, by keeping it simple, and by minimizing the dependence of complex or non-universal OS platform features. We built our own platform-independant APIs and ensured that they did what they were supposed to do on all platforms. And we made sure that the APIs were easy for developers to use. We paid attention to the order, the types and the naming of parameters.

" alt="" src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif" data-mce-src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif"

Having had years of experience in compiler design, database design, operating system design and configuration management design, we started with a number of design guidelines, such as not using goto's when coding, but more extensive than this. Things such as attention to how to design each function of an API and whether to have one function with a lot of options or several functions with few options, or both, and in what order the parameters should occur. Like using C pointers only in a very restricted manner to pass arguments and to reference entire records of internal or database data. Like being consistent across all of our software, especially with coding standards, and training new employees on the design rules in such a way that they could understand how each rule helped. Like establishing good naming standards that helped code readability rather than hindered it. Like instituting peer code reviews with a focus on design rules and refactoring as part of those reviews.

Some might be surprised that on the C-language side of things, we actually significantly restricted the use of the language. C is too flexible. we replaced the "case ... break" with a "_case" macro, ensuring the previous case was ended before the new case. We disallowed loops where the conditions were at the end of the loop rather than at the beginning - simply because it makes the code harder to read. We severely restricted how pointers could be used. Of course we eliminated "goto" and statement labels (other than for switch statements). We eliminated nested functions, simplifying scoping rules and increasing re-use. We assumed that #define's would eventually be replaced with real variables and so named the defined variables contrary to the usual upper case convention. We replaced C-Strings with our own descriptor-based strings so that strings would not have to be copied to point to sub-strings. And so forth - whatever we could do to reduce errors and improve readability and simplicity.

We introduced some basic guidelines. Simplicity is more important that optimization. Keep functions under 50 lines, generally. Mixed case for global names, including global
variables (with guidelines) and function and type names.  Don’t use abbreviations here, unless they were well defined acronyms (e.g., DB), and lower case for local names, including field names and local variables. As local names had full context within a few lines, shorter names were recommended to make the code more concise and hence easier to read. Common local variables, such as str for a general string, pos for position within a string, i, j, k for normal arbitrary loop index values, n and count for counts. We focused more on good names for boolean locals. Also, we made sure that the names chosen made the code (e.g., if statement) more readable. This is especially important with booleans, arrays, and function names.

We focused on constant code refactoring, because software architecture was everything. Copy and paste was replaced with factoring and this not only kept the code minimal, but eliminated the multiplication of bugs with each copy and paste operation. It further allowed us to improve the factored code rather than having to try to track down all of the copy/paste locations.

Getting the Requirements Right
Neuma's CM product evolved over a period of about 20 years (so far). So how did we get the requirements right, or did we? Well first of all, the wealth of large project CM experience helped. We knew, for example, that change packages (aka updates) had to be at the center of any solution. We also knew the value of seamlessly integrating applications and the derrived benefits, but it was still not easy. For one thing, GUIs were rare and evolving in their infancy 20 years back. Networking dealt with connecting mainframes together (except in the Unix world), and not users. And our CM tool experience was based on fixed company requirements, once for an IBM mainframe supporting hundreds of users, and once for a small network of DEC Vax mainframes supporting a couple hundred users. Keeping the command line interface simple was important. So was the architecture of each language we were using, in both cases proprietary.

The focused in-house requirements gave us a tremendous blind side, especially because Windows wasn't mainstream for development, and Unix was just expanding its foot-hold. There were no file system standards to adhere to (i.e., make the design architecture mimic file system architecture). As a result, our first releases of CM+ focused on a complex folder/module/section paradigm, where each module, which shared a common base name, was composed of several sections, identified by the file suffix. For example a C-module had a .h and a .c component, and in our case a .x component as we preferred (and still do) to keep externals separate from all other header definitions. An Oracle form had a different set of sections. An assembler language module had a .inc and a .asm component.

Although the product let you define your own module types (called groups), and this was a selling feature for some shops, we soon realized that every shop worked differently.  It wasn't always easy to package things into neat module groups, especially because the groupings overlapped. As a result, it became nearly impossible to automatically load in a new project. Even if all of the module groups were defined up front, when the system encountered an overlapping section (i.e., one that could be part of several different groups), it didn't know which group to assign to it.

As well, in the old days, apart from Unix, executables were often built from all of the files in one directory. There was no overlapping name space and it was easy to go from a file name to exactly what executable, or executables if it were shared, it belonged to. This flat name space made things very easy and a few of the older CM tools adopted it. But in the end, as the Hierarchical File System took precedence and users wanted the same name in different products, and, especially with O-O Design, the same name in different subsystems of the same product, we had to admit that our design was inadequate.

" alt="" src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif" data-mce-src="/sites/default/modules/smart_paging/plugins/wysiwyg/smart_paging/images/spacer.gif"

Our first attempt to fix the problem was to allow a flat name space per product. But this was inadequate. This resulted, in the mid-1990's, in Neuma having to, not only completely re-do its product file-handling architecture, but also in having to improve its context specification ability. In a flat name space, some aspects of context aren't as important from a file management perspective. In a hierarchical, product-based, overlapping name space, it was crucial. Furthermore, through all of this, we had to ensure that our existing customers would have the option of continuing with the flat name space or moving to an overlapping name space.

The point is, it was not easy to get the requirements right. And requirements continue to evolve. So what's the solution?

Solution Architecture
One of the main reasons we were able to weather the storm is that we focused on architecture at different levels. We did not need to know what the CM requirements were to understand what a NG database must look like to support a general set of engineering applications.

On top of that, we knew from the start, that automation and zero administration were important goals. Even after completing the initial NGDB architecture, we took the time to understand what potential clients said was number one in our target market requirements (customization) making the tool work the way the customer wanted. This molded most of our efforts beyond the NGDB design. We would seriously consider whether or not customization would be required for each feature and err on the "yes" side. But we would also consider how to build an architecture that was easy to customize.

When GUIs came along, this became a priority as well. If every site had different customizations, we did not want to get into customers having to paint forms, create dialogs, etc. We wanted the tools to do the tedious work, while the customer just identified what information was desired.

In fact, with each release, one of the largest components of the release is to support customization more easily and more widely. If it's easier for the customer, it's easier for us to sell and to support. So the business case for this effort is easy.

At the same time, we would not compromise on reliability. This meant simplicity where possible, especially when interfacing to outside elements. A multiple site solution has to interface with outside elements so must be kept simple if automation is to result. An automatic baseline capability is anything but simple, by definition, but does not have to interface to outside elements, as long as all of the information is in the CM repository.

Aim High
It's complex, and yes, gut-wrenching, to bite off more than you can easily handle. But if you don't bite off enough, you pay for it later. The single biggest problem with the software industry, from a solution perspective, is that it rarely aims high enough, with noteable exceptions. It says, this what we know how to do so we'll provide this bit of functionality. Eventually the architecture gets maxed out.

Instead, identify what the solution needs are: like zero-administration, small footprint, etc., and make your architecture stick to it. In the end it will pay for itself. People frequently ask, "How can you fit so much in so small a package?” It's because we aimed to support a full ALM solution from the start. We didn't have to build 10 solutions and then add additional glue to partially integrate them. We identified from the start that an ALM solution was necessary, and that we didn't understand where the boundaries of ALM would eventually end up.

If you hire different people to build the different walls of your house, it will cost more, there will be more glue, it will take longer and your overall structural integrity will suffer. Identify the common properties and make the tools to build all of the walls the same (dry wall, aluminum beams, insulation type, processes, etc.). Then it's easy to look at the structural integrity up front.

Don't try to build reliability into a software product after all the pieces have been built, as it won't happen. The increase in complexity will negate any attempts at reliability. So, too, with the other attributes: We want the product to be small-IT in the end, not big IT.

What About CM
So there are some lessons in the CM world - CM vendors take note. These are simply basic requirements. The ALM applications have to work together. We want easy-to-use low-training applications. Zero administration. Small footprint. Easy branching. Easy baseline definition. Change-based management. Support for development streams, and in general support more for the normal way of doing things (make that easier than the exceptions). Support for multiple site operation. Easy backups. High reliability and availability. Easy to customize, extensively if necessary, while eliminating the necessity as much as possible. And so forth.

It's not sufficient to look at a piece of the solution and apply good architecture to each piece. It has to be a solution-wide effort. Making multiple site operation work for the files in the solution does not give me a Multiple Site ALM solution. Consistent backups for the file repository do not give me consistent backups for the entire ALM repository. If even one piece has high administration, the solution will appear to have high administration. The same goes for reliability: the weakest link will be the most noticeable.

The first two generation of CM tools did not abide by these lessons. The next generation tool must, or it won't be considered a third generation tool. The good architecture of a tool will be much more apparent in third and fourth generation CM tools:  at cost, risk, roll-out time, resource requirements, reliability, and accessibility levels, to name a few.

To Sum Up
If you want to build better applications and applications that are more reliable:

·         Understand your requirements well, and expect them to change

·         Aim high, not only to meet, but to exceed your requirements

·         Put Architecture first, at the solution level and at the design level, using experienced architects

·         Generate re-usable components that will grow in reliability over time, and learn how to build APIs that are widely re-useable

·         Train your development team on your architectural requirements and guidelines, and do peer reviews against your guidelines

·         Understand the processes surrounding your application, and improve on them

Yes, you'll still have to document your product requirements well, especially so that you can verify them against the candidate releases. You'll also be able to better withstand the storms of changing requirements, evolving standards and competition.

I'm sure I've only scratched the surface, but maybe I've also ruffled a few feathers.  Let's hear from you.