Progress / OpenEdge ABL Language and DBMS Considered Harmful

24 minute read

I won’t go into any detail about how I got the job I currently have programming in Progress OpenEdge ABL, but suffice to say I knew I was in deep trouble from the start when most of the documents about the Progress language I read when learning the language were more focused on “business logic” and how the language makes things “easy” on the programmer rather than displaying the language’s real raw technical strengths.

The cold chill down my spine that I initially felt upon glancing at the language documents has since turned into a fiery rage that burns with the intensity of a thousand suns after having learned it. I have done my best to make do with the language’s limitations, and I believe I have pushed it to about as far as it can go without directly linking into external .NET libraries.

I am aware that Progress (later Progress 4GL, now OpenEdge ABL) was really created as a DSL to make business data management and reporting simple, and that a lot of the criticism below is harsh when comparing it to a “real” programming language. However, the fact that Progress corporation marketed it as a 4GL wizz-bang language that could do it all instead of the DSL that it really is, means that it is exempt from such consideration (they preferred to knock down the “3GL” languages I am contrasting it with below as being less capable!). This is especially true considering that any businesses that bought into this hype and managed to become successful are now stuck with an antiquated, closed language and database management system, making it impossible to keep up with the latest advances in technology unless Progress corporation enhances the language to add features that are not bolted in to the system (which, of course, costs money to upgrade to version n+1).

I see no reason for a business that needs data management and reporting to use OpenEdge ABL over any of the well-known (and free) programming languages and database management systems in the year 2010. The power of all the external libraries that exist in these languages simply makes it a no-brainer, and the ability to drop down into something lower-level and more powerful means that you are not constrained by a domain-specific language like with OpenEdge ABL.

Finally, I am well aware of the negative opinion some have of the “considered harmful” type of essay. I initially tried to think of some positive attributes of the Progress OpenEdge ABL language and DBMS to include in order to make this a balanced “positives vs. negatives” essay, but I honestly couldn’t think of any.

So, without further ado is a list of grievances that I have with the language and database system that I hope will provide reasons for anyone considering using Progress OpenEdge ABL to reconsider!

DBMS: No inherent concept of inter-table relationships (i.e. FK).

Progress’ database has no concept of inter-table relationships! I really can’t understate how huge of a weakness I believe this is for a modern relational database - I would go as far as saying you cannot consider Progress a modern relational database for this reason.

Consequently, this has major repercussions for the ABL language, whose strength is its simple integration with database access, because even though doing simple queries is relatively easy, the “root of the tree” so to speak is rotten, which kind of makes the fact that you can easily get to the branches a moot point.

For a database administrator trying to maintain meta information about a very large database with thousands of tables becomes a Herculean task, because there is no help from the database internally. Any table relationship information must be stored and maintained externally from the database, which adds overhead and increases the possibility of creating faulty information.

I will mention that in the ABL language there is a querying shortcut, however, where if you use identical field names (and at least one unique index) in differing tables you can do

FOR EACH order-line OF order

However, do not mistake this for an underlying concept of inter-table relationships. This is nothing more than syntactic sugar for doing simple table joins.

ABL: Division of integers isn’t “integer division” (it rounds up!).


i = 1 / 2.
MESSAGE i. /* Displays '1'! */

Apparently, all types of division are implicitly done as DECIMAL, even if you are dividing two INTEGER datatypes and storing the result in another INTEGER. Therefore, in the example above, the division of 1/2 results in a DECIMAL value of 0.5, which, when assigned to the INTEGER datatype, is rounded up to 1 (!). The workaround is to do this:

TRUNCATE(1 / 2, 0). /* Force integer division (returns '0') */

ABL: Arrays are 1-based rather than 0-based.

Some might call this a religious argument, but there is a reason that most programmers start counting from zero. I’ll let Edsgar Dijkstra take this one.

ABL: Inability to declare a variable as constant.

Using preprocessor directives in many cases can achieve the same result, however, experience from industry long ago has shown the preprocessor method tends to be more error-prone (similar to the citation Dijkstra uses in the above 1-based vs 0-based array link). See C Elements of Style, chapter 6 for some examples.

I would also add that it also makes enumerations in object-oriented languages much better, because a defined constant obeys scoping rules. For example, you can create a class called Color and define the color RED as an integer enumeration equal to 0xFF0000; if you define the color RED as a public constant you can then have other classes refer to it like so: Color::RED, and they will simply get the value 0xFF0000. Pretty slick!

Update 2010-10-02: Just wanted to point out that the above Enum example can be achieved like so: ```openedge CLASS Color: DEFINE PUBLIC STATIC PROPERTY RED AS INTEGER INITIAL 16711680 NO-UNDO GET. PRIVATE SET. END CLASS. ``` I recently used a similar method for the Stomp client framework. I am still hoping for a const keyword for more general language use, though.

ABL: Primitive variables are not very primitive.

My definition of primitive here is a value type datatype built into the language whose behavior cannot be extended by a program and the size of which is expected to be a fixed size at compile-time, e.g. INTEGER or LOGICAL. I don’t really think of LONGCHAR as a primitive type. According to ProKB #P144589 (my emphasis added):

As for data, such as CHARACTER, LONGCHAR, etc variables and parameters of any kind, we do not pre-allocate memory for them, except for the 'coordinates' of the variable or the parameter. ... The actual data itself is all variable length, whether in memory or in a record.

To me, that means that the actual stack memory allocated for variables are always pointers to variable-length heap memory. I can’t help but think that this can have serious performance issues due to the efficiency differences between stack and heap memory.

I also do not like the fact that anything (even “primitives”) can be assigned a value of ?, which most ABL programmers who come from other language backgrounds approximate to be ABL’s NULL pointer value. This leads to weird things like how a LOGICAL (ABL’s boolean datatype) is actually not boolean but tertiary - it can have the values TRUE, FALSE, or ?. This is actually how message boxes with “Yes” “No” and “Cancel” buttons in OpenEdge ABL work - the “Cancel” button assigns the LOGICAL variable to ?.

ABL: Un-optimized language compiler.

When I had to learn Progress for my job, I had been taught from an old manual that stated that when ASSIGNing to multiple variables sequentially, it is more efficient to group the assignments into a single ASSIGN statement, e.g.

ASSIGN i = 0
       j = 1.

is supposedly more efficient than

ASSIGN i = 0.
ASSIGN j = 1.

I recently was looking at the 10.1C documentation and was surprised to find that that information is still in there. I found it hard to believe this is still an issue, so I even did a test compile in 10.1C and sure enough, the resulting object code was larger for the two-ASSIGN statement scenario.

Such a simple thing as this should have been optimized away a long time ago by the compiler. The fact that the compiler cannot handle such a simple optimization leads me to believe that very little is being optimized by it at all. This leads to developers always having a faint voice in the back of their head saying “Will this compile into optimal object code?” which takes away time and effort from the development process.

This is one of those things that was an issue at least 10 or 15 years ago for programmers who code in modern languages. The fact that I even have to think about this in the year 2010, quite frankly, is a testament to how antiquated this language is and sends a message that Progress corporation has little interest in modernizing its core. I hope that that doesn’t come across as being too incendiary, but it is one of the issues that I feel very strongly about.

Just for the heck of it, I also checked to see if the compiler could handle a somewhat more complex optimization - converting tail recursive calls into iterative:

PROCEDURE tailcall:
    IF ipiNum EQ 99999 THEN
      RUN tailcall(INPUT ipiNum + 1).
RUN tailcall(INPUT 1).

On success, nothing should happen; on failure, the call stack will overflow (assuming 99999 procedure calls is enough to overflow the call stack).

The result, of course: the stack went “boom”.

ABL+DBMS: Un-optimized query compiler resultsets.

A while ago, I had found an interesting scenario where I had a query that looked like this:

FOR EACH table
  WHERE table.field1 EQ "foo"
  AND   (table.field2 EQ "bar" OR table.field2 EQ "baz")

that was strangely running slow. I was using an index, and the XREF did show that it was using the expected index. I changed the query to look like this:

FOR EACH table
  WHERE (table.field1 EQ "foo" AND table.field2 EQ "bar")
     OR (table.field1 EQ "foo" AND table.field2 EQ "baz")

And presto, the query ran much, much faster, even though mathematically the two different syntaxes are identical (the XREF listing they produced was identical, showing they were both using the correct index). I did spend a little bit of time looking at Progress’ knowledgebase, and supposedly certain phrasings can cause the resultset (how the DBMS returns results to the client session) to be broken up differently which is probably the cause of the problem.

However, like the above compiler complaint, this is the year 2010 and not something the programmer should have to worry about, especially given the simplicity of the above example and the marketing of Progress ABL as being a database querying powerhouse (the above example should rather be an example of how well the language performs in such a scenario). If I were doing 15 table joins I would cut it some slack.

ABL: Limited multiprogramming capabilities.

There is no support for forking or threading processes, which allow a much more lightweight approach (by keeping process access in a single-user context) than sockets, which is the only method that ABL allows. Granted, the typical Progress ABL application probably will not have a use for this for things like number crunching. However, forking/threading really shine in networking applications, which can certainly arise in a data-driven environment like Progress, where interaction with non-Progress databases may become important. I won’t get into fine-grained details for those who have never written a network application, but the ability to have a process to handle incoming requests that forks or threads a new process for each request is much simpler (and entirely different) to deal with than a socket-based approach. Not to mention, the number of sockets that can be opened is kernel-limited, which I think is generally around 65,536 (if memory serves).

I do remember reading about a library some people were working on to try and simulate threads using sockets, which is definitely a valiant effort, but I just don’t think you can achieve the same simplicity (and certainly nowhere near the same performance without raw language support to hook into the kernel) as a fork or thread-based approach.

ABL: Lack of useful operators.

Modern languages have operators such as ++, --, +=, -=, *=, /=, %=, etc, which allows for things like i++ as a shortcut for i = i + 1 or i += 7 as a shortcut for i = i + 7 or cMyString += "foo bar baz" instead of cMyString = cMyString + "foo bar baz".

This can get really annoying when your standards dictate using the ASSIGN statement, e.g.

ASSIGN cMyString = cMyString + "foo bar baz".

instead of

cMyString += "foo bar baz".

ABL: Lack of bitwise operators and bit-shifting capability.

A couple months ago I had to implement an object’s hashcode; I had planned on hashing a few distinct private variables and then combining each hashcode together using either a bitwise-and or an exclusive-or, e.g.

hashcode(obj) = hashcode(obj.var1) ^ hashcode(obj.var2) ^ hashcode(obj.var3)

Since Progress doesn’t support it, however, I had to do something really boneheaded, like

hashcode(string(hashcode(obj.var1)) + string(hashcode(obj.var2)) + string(hashcode(obj.var3)))

If it was something important, I probably would have worked out something more robust, but it got the job done.

Also, although not relevant anymore (due to pipelined processor architectures; thanks Wikipedia), you can’t do fun things like a XOR swap (where you don’t need to allocate a temporary variable for swapping primitives).

ABL: Inconsistent error handling.

Some errors use global ERROR-STATUS handle (trip ERROR and push error message(s) to GET-MESSAGE), some methods return a logical and put error message into ERROR-STATUS error queue (retrieved with GET-MESSAGE), some return a logical but don’t put error message into GET-MESSAGE (these are the worst - they push error messages to stderr and if you don’t do this:

ASSIGN lSuppressError = methodThrowsError("Foo") NO-ERROR.

then these will be visible to the user!).

ABL: MODULO using negative LH operator is undefined behavior.

This one is somewhat nit-picky, but it really bit me one time and was annoying because no warning was produced - it simply returned weird results. After checking the documentation, it turns out that there is a note “The expression must be greater than 0 for MODULO to return a correct value”.

It just seems very silly to me, because modulus is such a simple operation. It would be very easy to write a wrapper class around this that checks the left-hand operator, does a “positive” modulus if the operator is negative, and then re-applies the negativity after the operation. This really needs to be fixed and should have been a long time ago.

Test case:


ABL: No native multi-dimensional arrays, e.g. INTEGER[][]

This mostly disappointed me because the language used to be written in C and it would seem like an obvious thing to inherit. This isn’t the end of the world, though, because the language does support TEMP-TABLEs, and a similar construct can be created like so:

DEFINE TEMP-TABLE ttThreeDimensionalArray NO-UNDO

However, there is a reason that in C++ or Java one does not always use a Vector or ArrayList: there is more overhead because of the way the dynamic memory allocation works and the extra work that goes on behind the scenes to maintain it. Sometimes you just want a simple chunk of (preferably stack) memory with no overhead.

The good news is that the work ABL was created for very rarely needs such performance considerations.

ABL: Temp-tables/Workfiles do not support all datatypes

LONGCHAR or MEMPTR, specifically. See Progress Error 11406 “Temp-tables and workfiles do not support the LONGCHAR data type.” and Progress Error 3347 “You can define and access a MEMPTR data type only through program variables.”

This would have been useful for me in at least one case, where I wanted to store a TEMP-TABLE that contained a CHARACTER that represented a file name and a LONGCHAR that contained the file contents loaded into memory. The annoying workaround I use is to wrap these datatypes in an Object:

    File        : LongcharWrapper.cls
    Purpose     : Wraps a LONGCHAR variable in an Object
    Description : Used in situations where you need to store a LONGCHAR but
                  Progress can't seem to handle it (e.g. in a temp-table)
    Author(s)   : Abe Voelker

USING Progress.Lang.*.



  CONSTRUCTOR PUBLIC LongcharWrapper():
    SUPER ().

    SUPER ().
    ASSIGN lc = iplc.

    RETURN lc.

    ASSIGN lc = iplc.


ABL: Poor Unix batch mode error handling.

Added 2010-08-31

There are a couple issues that I have with ABL’s batch mode error handling when being ran from Unix. The first is that the return code for some programs that should be “error” actually returns “success”. For example, if I run the following OpenEdge ABL code (which clearly produces an infinite loop that runs out of stack memory):

RUN p.

  RUN p.

using the following shell script:

umask 000

if (mpro -b -p stackboom.p > stdout.txt 2> stderr.txt); then
  echo "success"
  echo "fail"

will output success to the shell, even though the output stream is filled with error messages:

WARNING: -nb exceeded. Automatically increasing from 90 to 122. (5407)
WARNING: -nb exceeded. Automatically increasing from 122 to 154. (5407)
WARNING: -nb exceeded. Automatically increasing from 154 to 186. (5407)
WARNING: -nb exceeded. Automatically increasing from 186 to 218. (5407)
SYSTEM ERROR: -s exceeded. Raising STOP condition and attempting to write stack trace to file 'procore'. Consider increasing -s startup parameter. (5635)

The second issue I have is that batch mode ABL programs do not write error output to stderr. Everything goes to stdout. I had initially assumed that any Progress error would go to stderr, so I tested this using the following code:

UNDO, THROW NEW Progress.Lang.AppError("FooError", 1337).

which did produce output, however, the output went to stdout. Upon further investigation, it turns out that all error messages that are produced (even system errors) go to stdout. My guess would be it’s another feature of Progress that tries to oversimplify things in an ill-guided attempt to make it “easier” on the programmer. Unfortunately, this type of thing leads to poor program design, which leads to enormous amounts of technical debt.

OOABL: Objects are enormous

I have hit memory allocation of over 1GiB just by creating a couple dozen objects with small text files (total appended size not greater than 1MiB) loaded into memory as LONGCHARs. This is a big issue because loading a text file into a LONGCHAR is about the only way to load a file into memory, short of doing a lot of unnecessary work with MEMPTRs. I don’t know what else to say other than there is just way too much memory allocated in the above scenario; I’m not sure if this is strictly to do with the LONGCHAR datatype or if it is an object-oriented ABL issue.

Dealing with memory in general in object-oriented ABL is annoying because there is no fine-grained control of the amount of memory to be initialized, nor is there any way to tell how much memory a class instance will take up when initialized after completing the class structure (.cls file). In C++ the sizeof operator provides this functionality; in Java it is not so simple, but there is a package that offers the needed functionality - java.lang.instrument.Instrumentation.

The way I have been measuring object size is to start up a session, record the starting process memory allocation, run a small program that initializes a large number of the objects to be measured - say 10,000 or so - and subtract the starting memory allocation size from the current size (divided by 10,000 of course). This is far from an accurate estimate, but it can give a ballpark figure for the size of an object.

From my own estimates, an Object with only a CHARACTER variable (and no non-inherited methods) appears to require approximately 21.6 KiB! This is much higher than I would have ever expected; it makes me think of a CHARACTER variable more as allocation for a medium sized text file than as a replacement for a ‘string’ datatype like C++/Java offer, of which I think most programmers expect the starting allocation to typically be for a paragraph of text or less.

OOABL: Manual memory management (no garbage collection).

Update: This was implemented in OpenEdge 10.2A

This is from the “Getting Started: Object-oriented Programming: Getting Started with Classes, Interfaces, and Objects” manual, subsection “Managing the object life-cycle” (my emphasis added):

You are responsible for the lifetime of a class instance. You must delete the object when it is no longer needed. If the variable used to hold the object reference obtained from the NEW phrase goes out of scope and the object has not been deleted, the object remains in memory even though there is no reference to the object. This can lead to a potential memory leak, as it can with any other dynamically created objects in ABL.

Therefore, as long as you need the object, you must also ensure that you maintain an object reference to that object. You can always assign the object reference to another variable before it goes out of scope or pass it to another procedure, where you can continue to manage the object until you finally delete it.

Deleting a class instance

To avoid memory leaks during the execution of the application, you must delete an instantiated object when it is no longer needed. To delete a class instance, you must use the DELETE OBJECT statement.

I do not understand this reasoning in the slightest; words cannot express how vehemently I loathe this whole concept. The fact that OpenEdge ABL does not give you fine-grained control of memory allocation is understandable because it is a domain-specific language that tries to make things simple for the programmer. However, to then go on and expect programmers to always have to free their own memory (otherwise creating memory leaks) is plain stupidity. This is simply the worst case scenario of memory allocation that mixes the weaknesses of Java and C/C++’s different memory allocation methodologies.

OOABL: No support for abstract methods or classes.

Update: This was implemented in OpenEdge 10.2B

Interfaces are nice, but sometimes I want to implement some behavior in the superclass and stub out some behavior for the subclass to implement! This also makes implementing some design patterns, such as the Template Method pattern, impossible. This is a very serious problem for those trying to make well-structured object-oriented code!

OOABL: No interface inheritance support (sub/superinterfaces).

Update: Progress claims this is planned for OpenEdge 11.

There is no doubt that when ABL implemented object orientation it mostly copied Java’s methodology. What separates Java from C++ when it comes to OO is that C++ supports multiple inheritance of classes, while Java decided to accomplish the same concept by supporting multiple inheritance of interfaces. Therefore, in Java, a lot of design gets pushed to interfaces. Interfaces, like classes, need to be able to support code reuse. Unfortunately, ABL sucks in this regard because it doesn’t allow interfaces to inherit from eachother, so it defeats a lot of their utility.

To see an example of good design using interface inheritance, check out the how Java’s java.util.Collection extends java.lang.Iterable.

Here’s a test case for this behavior:


Which currently (10.2B) returns a Progress error, stopping you dead in your tracks:

Inheritance is not supported for interfaces. (13046)

OOABL: Cannot use sockets within a windowed environment.

The problem is that you cannot handle socket reads internally to a class; you must spawn an external persistent procedure to handle them. To then be able to actually read the data from the socket, you must do some type of input blocking in the class, such as WAIT-FOR READ-RESPONSE OF hSocket. If you execute that line within a windowed environment, you will get Progress error

  1. Batch mode processes work fine, however.

OOABL: WAIT-FOR fails inside of methods that do not return VOID.

This is annoying when you are dealing with sockets. The workaround is to put the WAIT-FOR statement inside its own method that returns VOID, then call that method. If the workaround is so simple, I’m not sure why the error needs to be thrown in the first place. See Progress error 5622 for more info.

OOABL: Cannot define a temp-table inside of a method.

The workaround for this is to put the TEMP-TABLE definition in the main class definition when possible. However, this will expand the memory size of the object, and is not always a plausible solution in general. For instance, I had a situation where I needed my method to make recursive calls to itself, which meant that I really needed a TEMP-TABLE that was scoped to each METHOD call’s block. Otherwise, each recursive call overrides the class global TEMP-TABLE! I had to re-do the algorithm iteratively just because of this weakness.

Update 2010-10-02: Brad Williams pointed out another workaround - use a dynamic TEMP-TABLE. This will, of course, introduce run-time inefficiencies since the TEMP-TABLE definition will not be considered until it is needed (no pre-compiling), and also the developer must be sure to clean up after him/herself or a memory leak will be created.

OOABL: CATCH/THROW is not well-implemented.

It does not work as one familiar with the Java/C++ paradigm would expect. For example, it would be nice if there were a way to THROW an error from a class and enforce caller CATCHing it at compile-time (in fact there is no TRY statement at all…).

I’m betting the difficulty is that OpenEdge doesn’t resolve these things at compile-time as it is. Probably that and there is a lot of backward-compatibility crap to be maintained. With that in mind, it would have been nice if they used different keywords entirely, because it isn’t nearly as useful as C++ or Java’s implementations of the same keywords.

It’s too bad, because the trip to the salt mines that is writing a library or framework in ABL could have been made a little easier if this was done properly.

OOABL: PUBLIC methods cannot access PRIVATE variables or methods using an object reference.

This one comes up a lot when I am writing a getter or setter for a class. For instance, if I had a class called Foo, with a private Baz object variable named objBaz and I wrote a getter like the following to provide a means to publicly access the Baz:


I will get a compiler error:

Cannot reference private member "objBaz" off of an object reference.

The problem seems to be the THIS-OBJECT reference, and the (silly) solution is to remove it:

  RETURN objBaz.

This compiles fine. However, it is somewhat common in my experience to use the same name for the locally-scoped input variable of a setter method as the name of the class variable, so I am in the habit of being explicit. In Java, this ambiguity is resolved by using the this keyword to explicitly use the class scope.

The inability of THIS-OBJECT to work as expected opens the door to potential gotcha’s when implementing setters:

CLASS BadSetter:


    ASSIGN cMyChar = cMyChar.

    RETURN cMyChar.

Another solution for the getter/setter-specific scenario is to use a PROPERTY instead of a private variable and plain public getter/setter methods.

Update 2010-08-24: Another example where this could cause headaches is with copy constructors. Consider this example:



  /* Copy constructor FAILS: */
    i = f:i.

Update 2010-10-02: Brad Williams has provided an example using a PROPERTY that provides correct behavior:

define public property i as integer no-undo get.
  private set.

However, I still think an object should be able to self-reference its private variables without needing to change methods to use a PROPERTY.

Well, that’s it for the list for now. I will attempt to keep it up-to-date by adding to and deleting from it as new issues are revealed to me and old ones are fixed. Feel free to comment, whether it be to rebut, add to the list, commiserate, or otherwise!