Adventures in Mapping Multivalues - Joins

What's a Join?

Relational databases are built on a mathematical principal called Set Theory.  To help us along, let's look at two sets. 

We'll call set A the PRODUCTS set.  Each element of the set contains the following properties:
PRODUCT_ID - always numeric, 1-4 digits
PRODUCT_NAME - text, up to 100 characters
COLOUR_ID - always numeric, 1-4 digits, may be empty

Note that not all products come in multiple colours, so the PRODUCT_COLOUR_CODE field may be empty in those cases. Most SQL implementations treat an empty value in a string field as a NULL. I explain what that means later.

We'll call set B the COLOURS set.  Each element of the set contains the following properties:
COLOUR_ID - always numeric, 1-4 digits
COLOUR_NAME - text, up to 100 characters

So let's look at the elements of our PRODUCTS set (I've shortened the column names to fit the blog):

P_ID P_NAME................. C_ID
   1 Widget                     1
   2 Gidget                     2
   3 Gadget                     1
   4 Thing-a-ma-bob             4
   5 What-cha-ma-call-it     NULL (not assigned)
   6 Whatever                   3

Here are the elements of our COLOURS set:

C_ID C_NAME..............
   1 Red
   2 Green
   3 Purple
   4 Blue

In the purest sense, a join is a super-set made up of combining two sets into one new set.  Let's look at the "simplest" example:

Full Cartesian Joins

The simplest join of all is a full Cartesian join.  While it's simple, it's very counter-intuitive for humans. Here is what a Cartesian join of the two sets above would look like:

P_ID P_NAME.............. P.C_ID C.C_ID C_NAME
   1 Widget                    1      1 Red
   1 Widget                    1      2 Green
   1 Widget                    1      3 Purple
   1 Widget                    1      4 Blue
   2 Gidget                    2      1 Red
   2 Gidget                    2      2 Green
   2 Gidget                    2      3 Purple
   2 Gidget                    2      4 Blue
...

Wait!!! What's happening here? Well, if you don't constrain a join, it will create a new set in which every member of set A gets paired up with every member of set B. In our case above, the result is 24 rows.  As stated earlier, while the logic is very simple, it's very counter-intuitive. Note that I've added the prefix P. or C. to the front of the C_ID columns to distinguish which one came from the PRODUCTS or COLOURS sets.

While the is the default processing, it is almost never what you want to do in real commercial systems. In SQL terms you would get this from the following syntax:

SELECT * FROM PRODUCTS, COLOURS

What this says in SQL terms is "Create a new set (the result set), containing all properties and all combinations of elements, without constraint, of the PRODUCTS and COLOURS sets".

Inner Joins

As a human being, I look at the two sample sets (table) above, and realize immediately, that I only want to combine elements from the COLOURS set in my superset when the COLOUR_ID property in the PRODUCTS set matches the COLOUR_ID property in the COLOURS set.  This is called an Inner Join.  There are two ways to accomplish this with SQL. The following two SQL commands product the identical result set:

SELECT * FROM PRODUCTS P, COLOURS C WHERE P.COLOUR_ID = C.COLOUR_ID

SELECT * FROM PRODUCTS P INNER JOIN COLOURS C ON P.COLOUR_ID = C.COLOUR_ID

The key thing is that we are constraining the result set to be only those combinations with matching COLOUR_ID properties in both sets.

For our example this produces the following result set:

P_ID P_NAME............... P.C_ID C.C_ID C_NAME
   1 Widget                     1      1 Red
   2 Gidget                     2      2 Green
   3 Gadget                     1      1 Red
   4 Thing-a-ma-bob             4      4 Blue
   6 Whatever                   3      3 Purple

Hmmm.  What happened to the element of the PRODUCTS set that had PRODUCT_ID 5?  Well, given our join condition, there was no element of the COLOURS set whose COLOUR_ID property matched the empty COLOUR_ID value of the PRODUCTS table element, so it didn't fit our criteria and got dropped. While we're getting closer, that's still not quite what we were looking for!  This brings us to...

Left Outer Joins

What you really want is this syntax:

SELECT * FROM PRODUCTS P LEFT OUTER JOIN COLOURS C ON P.COLOUR_ID = C.COLOUR_ID

What this is saying is that, reading your sets (tables) from left to right in the SQL command, the result set should always include all elements of the left-most set (in this case PRODUCTS), even if there isn't a matching element in the sets to the right (COLOURS). In this case you get this result set:

P_ID P_NAME............... P.C_ID C.C_ID C_NAME
   1 Widget                     1      1 Red
   2 Gidget                     2      2 Green
   3 Gadget                     1      1 Red
   4 Thing-a-ma-bob             4      4 Blue
   5 What-cha-ma-call-it
   6 Whatever                   3      3 Purple

I'm going to display it as most raw SQL outputs would display it:

P_ID P_NAME............... P.C_ID C.C_ID C_NAME
   1 Widget                     1      1 Red
   2 Gidget                     2      2 Green
   3 Gadget                     1      1 Red
   4 Thing-a-ma-bob             4      4 Blue
   5 What-cha-ma-call-it    NULL  NULL   NULL
   6 Whatever                   3      3 Purple

NULLs are Nasty!

NULLs are not the same as an empty string.  They have nasty implications. Think of NULL as an "unassigned" or "unknown" value. 

One artefact of set theory is tertiary logic. Lets think of phone numbers. If I have a customer table where I track phone numbers, there are actually three possibilities for any customer.

1. They have a phone number and I have captured it
2. They don't have a phone number and I know that for sure
3. I have no idea if they have a phone number or not, and/or don't know what it is

Note the the syntax "NULL = NULL" technically returns false. I may not know your phone number and I may not know your best friend's phone number, so I use the NULL value to indicate that.  That does not make them the same. Unless you two are room-mates, the two numbers are almost certainly NOT the same. That's why NULL = NULL returns FALSE in most SQL databases.

In the case of our PRODUCTS table, it may be desirable to pull back all PRODUCTS that don't define colours.  In those cases, all those NULLs really define the same state. It's like the N/A value.  Doesn't apply and all "doesn't applies" are the same when it comes to colours.  NULL causes more problems than it solves, but because it's an important part of set theory, it became entrenched in all relational databases, for better or worse.

How PICK Handles This

In the Multivalued PICK world, we do not bother ourselves with set theory. It's nice if you want to apply mathematics to your data, but even there, the only math is really assists you with is set theory.

In the PICK world, you'd add a dictionary definition record to the dictionary of the PRODUCTS file and use a translate correlative code, "TCOLOURS;X;1;1" to pull the name of the COLOUR in, using the COLOUR_ID field on the PRODUCTS table. the "X" in that correlatives is a code that says, "If there's no item in the target table (COLOURS) that contains that id, return an empty string."  The result would look something like this:

P_ID P_NAME............... C_ID.. C_NAME
   1 Widget                     1 Red
   2 Gidget                     2 Green
   3 Gadget                     1 Red
   4 Thing-a-ma-bob             4 Blue
   5 What-cha-ma-call-it
   6 Whatever                   3 Purple

No NULL values exist in the PICK world. You would have to come up with a special value for "unknown" or "N/A" and handle it explicitly. That can be done and would be far more obvious and intuitive than SQL's NULL value.

Dick Pick and Don Nelson designed their database to store its data as a sparsely populated string, called a dynamic array.  Empty strings were quite common and very efficient. Empty strings were often called NULL, but did not have the same meaning as in SQL.

How Did Liberty ODBC Handle This?

PICK users were (and in some cases still are) a very unique breed. They were typically business people, with no computer science training, who found the PICK system so intuitive that they could create their own queries, and in many cases help design their own databases.  This sometimes caused problems because with their lack of computer science background, they'd do things that would cause their system to grind to a halt.

As we started moving these people into the ODBC world, they would try to pull data into Excel, and things would go sideways quickly.

One of the problems was that they would do a full Cartesian join, assuming that the SQL processor would be "smart enough" to figure out the obvious join conditions.  Our very tiny, simple example above produced 24 rows.  Let's imagine that you have your largest table, with 1 million rows of data, and you join it to a parts table with 10 thousand rows.  1 million times 10 thousand is 10 billion rows.  The result set will be 10 thousand times bigger than your biggest table. This likely results in two undesirable problems:

1. It will take forever to come back with results
2. You will likely run out of disk space and crash your system

We added logic to our SQL processor to detect a Cartesian join. If we saw one, we'd check a system parameter called "allow_full_cartesian_joins". The default value was FALSE.  A user could change the setting. If they did it using our admin tool, we'd put up a big warning and give the the option to cancel out or proceed.  This alone solved a lot of support problems.

The other thing we did was to ALWAYS teach our users what a LEFT OUTER JOIN was and why the ALMOST ALWAYS wanted it!

And lastly, we would train our users to understand what a SQL NULL was and what the pitfalls were, and that when they transferred an empty string to a SQL VARCHAR field, it would become a nasty NULL, and would often behave in ways that were counter-intuitive.

Adventures In Mapping Multivalues - Complex Data Parsing

Splitting Fields

A common situation in Pick is where you have a compound key in one file that links to another file.  As in the case of the control file, you may have a record with a key like C*RED.  Somewhere in your data, you may store that option value in a multivalue.  You may wish to display the fact that you have selected the colour "RED".  In order to do that, you could reference a dictionary that does a group extract on that field. The group extract "G1*1" says find the first "*" and extract 1 "*" delimited value. In our case it would give you the value "RED".

With this structure, our mapping tool would allow you to simply pick a dictionary that gave you the whole value for the foreign key, but also pick a dictionary that used the group extract to show you that extracted, or computed, value.

Sometimes the logic for doing this would get quite convoluted, and while I have seem massive correlatives in dictionaries, these are hard to maintain and very hard to create properly.  In these cases, our Select Method subroutines had a simple way of allowing you to do this type of processing.  To illustrate this, let's look at this example Pick record:

0001^18486^ABC Company^WIDGET]GIDGET^C*GOLD]C*RED\SZ*SMALL^2]1^
11595^22075

Note that everything up to the first attribute mark (^) is attribute 0 (attributes are zero-based), and the "]" signifies a value mark, and the "\" signifies a sub-value mark.

In order to process this option data, we want SQL to have a field that's a Colour field and a field that's a Size field. No such attribute exists in PICK, so what we could do is start by choosing two attribute positions that are not used by real data.  Let's say we look at the file and there are never more than 23 attributes, and no valid dictionaries reference anything higher than 23. We could leave some room for growth and pick attributes 31 and 32 as dummy attributes.

Next we'd create a dummy dictionary for attribute 31, called Colour and one for attribute 32 called Size. We'd use the comment field of the dictionary to note that it's used by the Liberty mapping for a computed value.  We need the dictionaries for the mapping wizard to create the SQL column mapping.

COLOUR
001 A
002 31
003 Colour
...
009 L
010 10
011 Liberty Mapping Virtual Field

SIZE
001 A
002 32
003 Size
...
009 L
010 10

011 Liberty Mapping Virtual Field

Then we'd create this subroutine:

SUBROUTINE SELECT.METHOD.INVOICE(ID, ITEM, PROCEED)
EQU VM TO CHAR(253)
EQU SVM TO CHAR(252)
* Run through multivalues - invoice lines
FOR M=1 TO DCOUNT(ITEM<4>,VM)
   * Run through subvalues in the current multivalue
   * these are the option values for that invoice line
   FOR S=1 TO DCOUNT(ITEM<4,M>,SVM)
      OPTION=ITEM<4,M,S>
      TYPE=FIELD(OPTION,"*",1)
      BEGIN CASE
      CASE TYPE EQ "C"
         * the multivalue position must match
         * but there will only be one Colour option
         ITEM<31,M>=FIELD(OPTION,"*",2)
      CASE TYPE EQ "SZ"
         * the multivalue position must match
         * but there will only be one Size option
         ITEM<32,M>=FIELD(OPTION,"*",2)
      END CASE
   NEXT S
NEXT M
RETURN
END

The SQL might look like this:

SELECT * FROM INVOICE_MASTER
Invoice_Id Invoice_date Customer...........................
      0001  11 Aug 2018 ABC Company

SELECT * FROM INVOICE_LINES
Invoice_Id Line_Id Product Quantity Price.... Colour. Size.
      0001       1 WIDGET         2    115.95 GOLD    NULL
      0001       2 GIDGET         1    220.75 RED     SMALL

The benefit of this is that you don't need a third view for the options, as in our hypothetical case we know (hypothetically) that any invoice line will only have one of any option type assigned to it.

Mixed Types

One of our customers that did deliveries had a very unique delivery field.  You could put a date and time in, or you could put text in.  The date might be 2018-09-12 15:30:00 or it might be "Tomorrow, any time before noon".  There was value in being able to sort any date fields that were entered as dates, but you also had to deal with text. We simply did a format check, and created two virtual fields. One would have the properly formatted dates and times, the other would have the text.

As you can see, Select Methods were an incredibly powerful tool for mapping multivalues.

Adventures in Mapping Multivalues - Data Cleansing on the Fly

In my last post, I talked about how we used a powerful feature that we alternately called Select Methods and simply Filters, to solve a problem with control files.  In this post, I'll take you through one of the powerful tricks that Select Methods let you play to control your data.

Select Methods, or Filters, are simply Pick/BASIC subroutines that get called for each record read from the file, before any processing is done on that record.  They have the following signature:

SUBROUTINE SUB-NAME(ID, ITEM, PROCEED)

It is generally recommended that you not change ID, but we won't stop you. ITEM is the item body, and as you will see, changing it is quite acceptable. In the last post we showed a way that you would use PROCEED to indicate if a particular record should be processed for the current view, or skipped.

Data Cleansing on the Fly

To understand the importance of this feature, you need to understand what we've mentioned in previous posts. You can define a dictionary definition for a file, if you wish, but this is merely suggestive. Your BASIC programs, and any number of TCL commands (Terminate Control Language was the command-line for Pick systems) could update any format of data into the file.  Examples are a programmer, fixing bad data, accidentally copies a fixed record into the wrong file, with the wrong format for that file.  A BASIC program with a bug, writes a text string into a numeric field. Or some kind of data corruption occurs and the data in a record becomes gibberish.

In the SQL world, you just can't do that, so we would crash a query if we encountered bad data.  We used to have customers insist that their data was good, so we would run our validator on their files and it would start kicking out reports of bad data. In all the years that I worked with Pick customers, I have NEVER run into one that did not have bad data in at least a few of their key files. In most cases, every file with more than 10,000 records had some bad data.

Customers had two options. One was to fix the bad data our validator found. This was great except that we would always ask how the bad data got there. The first time you got bad data again, the query would crash again.  In cases where data was being corrupted by a program or process unknown, there were a couple of simple fixes that could be done using Select Methods:

The first option would be to use a Select Method as a filter. Let's say that you have two fields, attributes 5 and 7, that need to be numeric.  The following program would detect if the data contained was not numeric and tell the program to skip it. As noted in the comments and the omitted code, you can also write a message with the item-id into a log file so you can fix it later.  This is a better option when you think the record is possibly complete garbage and you really don't want to play with it until you've had a look.

Here is what that subroutine would look like:

SUBROUTINE VALIDATE.MYFILEDATA(ID, ITEM, PROCEED)
IF NOT(NUM(ITEM<5>)) OR NOT(NUM(ITEM<7>)) THEN
   PROCEED=0
   * Optionally add code here to write the ID to an error log
   * file so someone can review it
END ELSE
   PROCEED=1
END

The other option is to cleanse the data on the fly. This is what that might look like:

SUBROUTINE FIX.MYFILEDATA(ID, ITEM, PROCEED)
IF NOT(NUM(ITEM<5>)) THEN
   ITEM<5>=0
END
IF NOT(NUM(ITEM<7>)) THEN
      ITEM<7>=0
END

This option makes more sense if you know what attributes are getting the bad data, but have reasonable confidence that the other data in the record is good.  This latter approach was one of the most common uses of our Select Methods.

Adventures in Mapping Multivalues - Control Files

What Shape is Your Data?

In a SQL world, a specific column in a specific table has exactly one data type and it cannot change.  In a pinch, you can represent everything with a varchar, but that's not very useful. Arithmetic gets tricky for numbers, and dates become impossible to sort!

Pick, on the other hand, much like XML, is essentially free-form.  You can define dictionaries for your data, but there is nothing that forces you to follow the dictionary layout, or even warn you if you depart, and in many Pick dictionaries, you don't have to look far to find two dictionary records that point at the same attribute, but have completely different definitions.

Just about every Pick system I ever worked on had a control file. About 75 to 80% of the time, it was called "CF".  The item-id of this file would normally have a structure like this:

Control-record-type "*" Identifier

The asterix was probably the most common delimiter in early Pick systems and carries through to many applications.  So some examples might be:

SIZE*SMALL
COLOUR*RED
INDUSTRY.CODE*312000

The thing with this file is that each record type would have a unique structure. For SIZE the first attribute might be an integer, which is the number of inches.  COLOUR might have 3 integers for red, green and blue, and INDUSTRY.CODE might have a text name of the industry, followed by some tax rates.

In order to process this and a whole bunch of other problems, we came up with a strategy called a "Select Method" or "Filter".  This was a subroutine that you would associate with a table that you defined.  This filter was a Pick BASIC subroutine with this signature:

SUBROUTINE SUB-NAME(ID, ITEM, PROCEED)

As soon as our SQL engine read a record for processing, it called this subroutine, passing the item-id in ID, and the full dynamic array text of the item body in ITEM.  You got to set PROCEED to 1 or 0 (true or false). If you set it to false, we'd skip the record.

So, for our example above, we'd define 3 table views on the CF file. The first might be called CF_SIZE and the filter method would look like this:

SUBROUTINE LF.CF.SIZE(ID, ITEM, PROCEED)
IF FIELD(ID,"*",1) EQ "SIZE" THEN PROCEED=1 ELSE PROCEED=0
RETURN
END

It would only process records that started with "SIZE*".

You'd then create another view called CF_COLOUR with this filter:

SUBROUTINE LF.CF.SIZE(ID, ITEM, PROCEED)
IF FIELD(ID,"*",1) EQ "COLOUR" THEN PROCEED=1 ELSE PROCEED=0
RETURN
END

This would only process records that started with "COLOUR*".

And you'd wind up with one called CF_INDUSTRY_CODE with this filter method:

SUBROUTINE LF.CF.SIZE(ID, ITEM, PROCEED)
IF FIELD(ID,"*",1) EQ "INDUSTRY.CODE" THEN PROCEED=1 ELSE PROCEED=0
RETURN
END

This would only process records that started with "INDUSTRY.CODE".

We would then define a different mapping for each record type.

Note that we avoided periods in table names as periods were significant separators in the SQL world. If you used them you'd have to quote your table names. In the early days, many ODBC clients wouldn't do this, so you were asking for trouble if you used them!

In addition to Row Select Methods, there were Row Delete Methods, Row Insert Methods and Row Update Methods that you could use to tweak how you wrote data back.  We'll deal with this in another post. Our next few posts will look at some more magic that we did with Row Select Methods, otherwise known as Filter Methods.

Adventures in Mapping Multivalues - Two Common Approaches

The Problem

In a previous post, I talked about what a Pick database was.  Pick had a way of representing complex data that was intuitive and matched the structure of common documents that it was modeling very closely. It was a very human-friendly way to represent that data. Note that this is much the same argument for XML data (that's for another blog post...)

In this post, I'm going to talk about how this might look in a relational database, then I'll talk about two different approaches that Pick uses to represent this kind of data:

• Correlated Multivalues
• Subfiles

Our Document Data

So, let's consider the example we gave of our Invoice. Here is how it might look in a printed document, with no consideration of how it actually is stored in the database:

Invoice 0001 Date 11 Aug 2018 Customer ABC Company

Product Options Quantity   Price............ Extended...
WIDGET  GOLD           2            $115.95    $231.90
GIDGET  RED            1            $220.75    $220.75
        SMALL
========================================================
Subtotal:                                      $452.65

Note that you have an invoice number, which should be a unique key, a date, and customer, that are all fields that have a single value.  Then you have a bunch of fields that have more than one value: Product, Options, Quantity, Price. We also have two calculated fields, one being a subtotal, and the other being the extended price (Quantity times Price), which has multiple values, one for each product line.  Note also that the Options value can have multiple values for each product value, so this is a sub-multivalued field, or subvalued for short.

Relational (SQL) Model

Here is one possible relational representation of that data:

SELECT * FROM INVOICE_MASTER
Invoice_Id Invoice_date Customer...........................
      0001  11 Aug 2018 ABC Company

SELECT * FROM INVOICE_LINES
Invoice_Id Line_Id Product Quantity Price....
      0001       1 WIDGET         2    115.95
      0001       2 GIDGET         1    220.75

SELECT * FROM INVOICE_OPTIONS
Invoice_Id Line_Id Option_Id Option
      0001       1         1 GOLD
      0002       1         1 RED
      0002       2         1 SMALL

Let's talk about some points here.  In the INVOICE_LINES table, we've added a Line_Id field, which combined with the Invoice_Id would be the primary key (unique identifier). If you knew for sure that you would never repeat a product in another line of your invoice, that any product would be unique within any invoice, you could skip that Line_Id field and use the Product value with the Invoice_Id as the primary key. That assumption would probably bite you. (In a multivalued database, you never have to create artificial keys like this.)  The same thing happens twice over in the INVOICE_OPTIONS table.  This mess of decisions, keys, and additional values creates code entropy, as programmers are required to do multiple reads, then navigate multiple data sets to process the data. One is generally safer to create a new key in these cases, but then you have a task of managing these keys when you write the data back, or insert (or delete) new invoice lines or options.

Correlated Multivalues

Here is the most common representation of the above in Pick:

0001^18486^ABC Company^WIDGET]GIDGET^C*GOLD]C*RED\SZ*SMALL^2]1^
11595^22075

Notice that in this structure, you have minimal wasted space. It was developed when memory and disk were both expensive, so keeping things small was beneficial.  The up-arrow character '^' is the attribute, or field delimiter. The square bracket ']' is a multivalue delimiter, and the backslash '\' is the subvalue delimiter. The actual delimiters are high-ASCII characters that were not commonly used in data in typical ASCII systems.

The first thing you see is what Pick calls the Item-Id. In Relational terms, this is a single field that is always the entire primary key. It is a unique identifier that is used to determine which bucket of the hash-file to put the data  in, or to find it in, if you are looking data up.

The next two fields are the invoice date (number of days since Dec 31, 1967) and the customer name. They both have just one value in them.

The next 4 fields contain multivalues. The first value of each of these fields is the data for line 1 of the invoice. The second value of each of these fields is the data for line 2 of the invoice.  For the Options field, the first value contains just that value. The second value contains two subvalues. The first one is the first option line for the second invoice line, the second is the second option line for the second invoice line.

For a Pick BASIC programmer who has read this record into a variable called X, he would access the first invoice line's product with this code:

PRODUCT.ID=X<3,1>

The second line's product value would be this:

PRODUCT.ID=X<3,2>

The Item-Id is not part of what comes back from a read (you need it to do the read), so it's often referred to as attribute 0.

To get the second option of the second invoice line, you would use this:

OPTION.CODE=X<4,2,2>

You basically had to know the attribute numbers for your fields and you could get at anything. One read statement got you everything for your document. Your record really was your document.

Subfiles

A less common, but still often used approach to representing this data is called a subfile. In a subfile, all the detail line data is contained in a single field. Here is how it would be represented:

0001^18486^ABC Company^WIDGET\C*GOLD\2\11595]GIDGET\C*RED,
SZ*SMALL\1\22075

All of your detail lines are contained in attribute 3. The first multivalue is all of invoice line 1's data. The second multivalue is invoice line 2's data.  The first subvalue of each multivalue is always the product, the second subvalue is the options, the third is the quantity and the fourth subvalue of each multivalue is the price.

Note that we ran out of delimiters, so we had to use a comma separator two delimit the second line's two option values.

So here is what a Pick BASIC programmer would do to get the first invoice line's product:

PRODUCT.ID=X<3,1,1>

You need to specify that you only want that first subvalue, or you'll get the whole invoice line. That may be handy in itself, but then you'd have to do a second line to get that data out.

The second line's product value would be this:

PRODUCT.ID=X<3,2,1>

To get the second option of the second invoice line, you would use this:

OPTION.CODE=FIELD(X<3,2,2>,",",2)

This line gets a bit more involved. The angle brackets only work with system delimiters, so it retrieves both option values, with the comma separator. The FIELD statement says to pull the 2nd comma delimited value out.

Again, your record is really the entire document.  Note also, that the BASIC implementation has data access and handling functionality built in. You don't have to call out to an API as it's all there!

Liberty Mapping Terminology

Liberty ODBC's SQL Mapping layer handled both of these structures very simply.

The Correlated Multivalues were termed a "Group Floating" table.  If you were dealing with Correlated Subvalues, you had a "Group Floating within Group Floating" table.

The subfiles were called a "Group Positional" table.  By using these structures, you would get the equivalent to the above.  You could define SQL columns that were the multivalue or subvalue offsets, effectively giving you the Line_Id and Option_Id fields.

When joining two tables where one was actually contained inside the other, our engine would optimize and read the Pick record only once for efficiency sake.

Upcoming

In some of my upcoming posts, I'll talk about some of the interesting and odd problems we had to deal with in order to efficiently map and process multivalued data in our SQL engine.

What is a PICK MultiValue Database?

Why I'm Blogging This

In some upcoming posts, I'm going to be talking about how we implemented an ANSI/92 compliant SQL engine and ODBC driver for this, and some of the interesting challenges we encountered. In some cases, I'll be talking about the relative merits of these systems vs. relational databases.  They all have their strengths.

In order to do this, I thought it would be good to have a bit of a primer about how Pick data is stored and processed. I'll probably reference this little blog post quite a bit.

It's Still Here

The first thing I want to say is that systems running on Pick based multivalue are ubiquitous and they are EVERYWHERE! In some cases they are legacy solutions that have been in place and working for ever.  Probably every fortune 500 company has at least one system, running in some department, that runs on it.  It's also still being sold as part of applications around the world.

Examples of industries and solutions that continue to run and/or sell best-of-breed solutions running on PICK include (but are not limited to) Automotive Retail, Libraries, Banking, Healthcare, Insurance, Manufacturing, Distribution, EMS, Hospitality, Municipal, Government, and many more.

Ignore them at your peril!

Where Did it Come From?

In about 1965, the US department of defense awarded a contract to TRW to create a system to manage Cheyenne helicopter parts.  Two engineers, Don Nelson and Dick Pick were assigned to the task, and started work on an IBM System/360 computer. They named their system GIRLS, which was an acronym for Government Inventory Relational Language System.

This was before commodity hardware was a thing. Somehow they got the idea to create a system based on a virtual machine that they could implement on any hardware.  There was some really good architecture and design that went into this system.

How Does PICK Data Work?

The database was a very simple concept.  Whatever document you needed to represent, you would have a data structure that contained all the information in one place.  With a single read, a programmer could pull back all the key data that was needed for an invoice.  In real practice, your invoice would reference a customer by an id, and the customer name would remain in the customer master file. Similar things would happen with part numbers and the part names.  The invoice would include both the header information and the detail lines for the parts being ordered.

The data is stored in a hashed file structure and each record is stored as a sparse string with delimiters separating fields, values and subvalues.  Let's look at an example.

Still working with the invoice concept, lets consider a manufacturing organization that sells three key products, widgets, gidgets and gadgets.  The invoice file has an invoice date, customer id, then it has 4 fields that contain multiple values (multivalues) in them.  These are product id, options, quantity and price.  When you display an invoice it might look like this:

Invoice 0001 Date 11 Aug 2018 Customer ABC - ABC Company

Product Options Quantity   Price............ Extended...
WIDGET  GOLD           2            $115.95    $231.90
GIDGET  RED            1            $220.75    $220.75
        SMALL
========================================================
Subtotal:                                      $452.65
Taxes
Totals

When looking at the document, you can see that you have header data that occurs just once per document, including the invoice number, date, and customer information. You would look up the customer's name from the customer master file.  You also have computed or calculated fields, like extended costs, subtotals, taxes and final totals.

The interesting thing to note is the invoice lines data. In a real invoice all the lines, an set of arbitrary size, is part and parcel of, and contained in, the invoice document.  For a single invoice you could have an arbitrary number of products being purchased. for each you would have a quantity and a price.  In our case, we further complicated it by allowing different products to have different options. You might choose a colour and a size as with the gidget we had on the second invoice line.  At this point we not only have multiple values, but one of the values (the OPTIONS value for the GIDGET line) has two subvalues, one for the colour and one for the size.

Don and Dick came up with a database structure that allowed that.  Your fields were delimited by field or attribute mark delimiters. Your multiple values were delimited by value mark delimiters and your subvalues were delimited by subvalue mark delimiters. The data for the above INVOICE file might look like this:

Note that the blocks are the field delimiters. the superscripted '2' is a value mark, and the superscripted 'n' is a subvalue mark. These are high-ASCII delimiters. Going left to right, we have the first field, the item-id (similar to a primary key) which is the invoice number '00001'. Then you have a date. This is represented as number of days since Dec 31, 1967.  Next we have the customer id 'ABC' for ABC Company. Then we get a field that has two values. This field and all the subsequent fields have two values, the first one is line 1 of the detail lines and the second is line 2 of the detail lines. This first multivalued field has the product ids of the products being bought.The next one has the options selected for each line. You can see that for detail line 1 there is only one option, but for the 2nd detail line (the GIDGET), we have two options: The colour RED and the size SMALL. Then we have the quantities of each of these and the unit price for each line. The price has an implied 2 decimals (it is storing the value as cents, not decimal dollars).

For a programmer, the invoice number gave you the whole document. You might need to do ancilliary lookups in the CUSTOMER, PRODUCTS and OPTIONS tables, but a single read and a single write gets you your data and saves it to the database.

In a relational database, the INVOICE file would need to be a minimum of 3 tables. The record represented above would take a minimum of 6 disk reads to get the same data. As you increase the number of detail lines and the number of product options, the complexity grows rapidly.  A programmer in this case needs to retrieve the data from the database and make sure he gets it all. Then he needs to properly relate the data sets in his program. The complexity is huge compared with the multivalued approach. 

So, in relational terms, you would probably have these tables:  INVOICE_HEADER, INVOICE_LINES, INVOICE_LINES, INVOICE_LINES_OPTIONS.

Think of how this would work if you stored your physical invoices like this. You'd go to the filing cabinet with header information and look up the header part of the invoice. Then you'd go to another filing cabinet where you'd have to find a separate page for each invoice line. You'd use the invoice number and line number. You'd keep looking until you couldn't find another line number for that invoice. Then you'd go to yet another filing cabinet, and for each invoice line for that invoice, you'd look for one or more pages that had information for the options (possibly multiple) for those invoice lines. I've seen invoices that had hundreds of lines in them. Can you imagine then having to organize and manage those hundreds of pages?  Instead of one invoice that was maybe 3 pages long?

There is also a performance component to this.  I remember working with a customer who had moved from a mainframe to a minicomputer platform, and had determined that they could not afford the disk head movement of a relational database.  They were a big customer of our ODBC technology as we enabled them to have a relation view of the data and use it with tools that required this, while still giving them the reduced disk head movement for their core application.

How Did You Access That Data?

Pick had a concept called a dictionary. You would have a file that contained field definition records associated with every file. It was completely optional, and once the BASIC language came out for Pick, it didn't enforce anything in it. In fact, you could have two dictionaries that defined the same attribute completely differently. Generally, one would be wrong and the other right, but I've seen where a single file has multiple different record types, based on a prefix in the key. Which dictionary you wanted to use depended on the record key!

You could create a dictionary definition item that actually used a value from the current record to read a value from another file. This was called a translate, and was very powerful.  For the example index above, you would take the customer id (let's say it's attribute 2) and use it to pull attribute 1 (customer name) from the CUSTOMERS file. That translate correlative would look like this:

001 A
002 2
003 CUSTOMER NAME
...
007 TCUSTOMERS;X;1;1
...

Similarly, you could use this to pull information from a PRODUCTS file.

When these systems were first commercially sold, salespeople would walk a machine in from a panel van (they were the size of a fridge, so this was quite a feat.)  Then they'd use it to create a couple of tables, use the editor to create some dictionary definitions. Create a script program using PROC and BATCH to enter data. Then they'd use a LIST command to print out the data.

So the command "LIST INVOICES" would be automatically expanded out to:
LIST INVOICES INVOICE.DATE CUSTOMER.ID PRODUCT.ID OPTIONS QUANTITY PRICE

which would give us this output:

It was not uncommon for this to be so impressive for the business users watching the demo that a cheque would be written and the computer left there. Then they had to get programmers in and write a system... but it was also not uncommon for a business person to dictate what the system had to do and create his own reports.  Compared with the options available at that time, this was a huge step forward, and many commercial systems that are still in use today had their start with a business owner looking over a programmer's shoulders. This was the first instance of agile and pair programming!

Later Add-ons

The first systems were written using a combination of PROC (a scripting language), BATCH (which was horrible, and impossible to make pretty - totally character based), and when you got stuck, Assembler.

Later, a BASIC implementation was added with built-in support for the multivalued database, and that allowed you to use extended features of an ASCII terminal.  These databases still exist and go by a number of names:

• D3
• Universe
• Unidata
• mvBase
• mvEnterprise
• jBase
• OpenInsight
• QM and OpenQM
• RealityX
• And others

Relational Value and Pushback

An interesting thing to note is that many applications are being written today to use what are called NoSQL databases. The rationale for these databases sounds like an ad for Pick.  While relational databases have their uses, there are clearly applications where they are just not the right solution.  

That said, a huge amount of investment has been made by companies like Oracle, IBM, Microsoft, Cognos, Business Objects and others into enhancements and tools that leverage relational databases. This has not been the case with Pick, so there are many applications where it is imperative that the regardless where your data originates, you need to get it into relational in order to leverage these tools and technologies.

The industry is clearly seeing a divide form between SQL and NoSQL databases, yet there is also growing clarity about when you should use each.  What's not as clear is that there is a NoSQL option that has already existed for a long time and that is widely in use, that's another option.

ODBC for PICK - Why You Shouldn't

Fundamental Mismatches

There were a number of very key, fundamental mismatches between PICK databases, and the ODBC model.  Enough in number and severity that many would say it's impossible, or "don't even try", but the group of pioneers that assailed this opportunity were not deterred by a few minor challenges.  Here are some of the things that we had to address.

Two Main Parts of ODBC

ODBC consisted of two main components. One was a SQL syntax, the other was an implementation of a Call Level Interface (CLI), which is much the same as an Application Programming Interface (API).  As an implementer of a driver, you could declare yourself as supporting one of three levels of each.  In our case we implemented the middle level of both.

Structured Query Language (SQL)

This was the biggest one, and I'll address it better in a separate post, but will touch on it here.  ODBC assumed SQL. While there were ways to bypass that, most tools that worked with ODBC would assume a certain level of SQL, and if you reported that you didn't support it, they'd just give up. So, if you wanted to support ODBC, you had no choice. You had to support SQL.  At the time that ODBC was released, while it was gaining popularity rapidly, both ISAM and Multivalued databases still outnumbered SQL databases.

SQL demanded certain norms, including that it was strongly typed and that it abhorred multivalues.  Multivalue databases like PICK and MUMPS (which was the other popular multivalued system, especially in health care) were, as the name implies, designed to work with multivalues. Again, I'll cover this in more detail elsewhere.  In short, the data models supported by each were very widely different.

Strong vs. Weak Data Typing

In PICK, everything is a string.  If anyone is familiar with XML, there are some definite similarities.  In XML you can have numbers (but they are simply strings containing numeric characters), and you can even have a schema that declares a certain XML element to be numeric, but in the end, if you don't enforce the schema in code, you can put what you want in that XML element. 

PICK had similar structures. In PICK, much like an XML schema, you had dictionaries. It was quite common over time, for PICK dictionaries to collect garbage. You'd have several conflicting dictionaries pointing at the same attribute (the PICK name for a field), and only one of them was really correct, or perhaps none of them were correct!  And it was not uncommon for a file's dictionary level to be empty.  Dictionaries were not enforced, but were useful for doing LIST, SORT, SELECT or SSELECT commands.  LIST and SORT were for creating user reports and SELECT or SSELECT were for activating a list of unique primary keys (called item-ids in PICK terminology.)

The CLI portion of ODBC, meanwhile, was designed for programming languages designed to produce machine code, and perform machine-level calls (hence "Call Level Interface").  This included Assembler, C, C++, and COBOL among other languages.  These languages were all strongly typed.  You would declare a variable as being a string (null terminated - no 16-bit Unicode in the early days - that came later) or an integer. The CLI had its own data types and would even declare the size of the storage for an integer.  The CLI had details of the sizing and precision of each data type, which you had to map to your machine's (or compiler's) data types.  This guaranteed that regardless whether your machine called a 32-bit integer a small int or a long int, you knew how big the data coming back from, or going into, the CLI had to be, and therefore could allocate enough space for it.  This allowed for interoperability between machine types, which was a key requirement for the standard.

Multi-threaded vs. Single User

PICK was a legacy system, originally built on the premise that a user was connecting to an expensive user license through a single, predefined serial port that ran through the building right to their terminal (or PC with a terminal emulator, like PK Harmony or Wintegrate.)  There was no multi-tasking and definitely no multi-threading in that system.

While ODBC didn't require you to support multi-threading, most applications using it expected to open multiple connections to do their work. Now, if you had two serial ports, you could open two connections, but this cost a lot, and required you to run another serial cable to your PC.  Even when we went to networked connections, the license cost for an extra PICK user was prohibitive, so this was another issue we needed to resolve.

Authentication and Authorization

In a pre-public-internet, pre-malicious-hacker age, we were very much aware of the security issues that we were likely to raise, and took these very seriously.  Although it was only a reality for a small number of users, network connectivity was already a reality for some, and we knew it was only a matter of time before it became the norm. We already had people dialing up systems over the telephone network, and I had seen people using dial-up, Prime-Net and then Internet (pre-commercial), and get connected to the wrong Prime Information system. This system was a NORAD system and had no password on the SYSPROG account (the PICK equivalent of root!) We logged off very quickly!

PICK's authentication and authorization model at that time was very weak and not designed for a broad network-connected world.

Performance

We were very much aware that we were likely to run into interesting performance problems for a number of reasons:

Data Architecture Mismatch

We were forcing round pegs into square holes. There were fundamental differences in how SQL and Multivalue databases architected their data and trying fool PICK into thinking it was SQL was likely to present some interesting challenges for performance.

Data transfer speed

When we started this, most PICK systems still used Serial I/O for their connectivity. There were solutions from Netware that allowed you to use the network, but the last bit of the connection was through a serial port concentrator, so you were still limited to the maximum speed of the serial I/O.  This was exacerbated by the fact that most PICK I/O channels designed for terminals would fall over if you burst data into them too quickly.

Challenge Accepted!

I've never found a challenge like this to be discouraging, but rather invigorating, and I've been blessed to work with many people who share that enthusiasm for solving difficult problems in elegant ways, so we got to work addressing the issues.  Challenge Accepted!

ODBC and SQL

Data Driver Difficulties

Round about 1990, my brother's friend John had a company called Paradigm that was writing drivers for Microsoft.  Structured Query Language (SQL) was becoming very popular, with Oracle the clear leader.  There was a problem, however.  It seems that every time you came up with a new application that needed to access data, you had to write a new driver for every database you wanted it to access. If you had, say, 5 applications that needed access to data and 5 databases that you wanted them to access, you had to write 25 drivers.  If the numbers became 10 and 10, you needed 100 drivers. This simply would not scale!

SQL Access Group (SAG) Call Level Interface (CLI) - Alias ODBC

Microsoft was not alone in struggling with this.In 1989 the X-Open Open Access Group had formed a specialized standards body called the SQL Access Group, with the initial members being Oracle Corporation, Informix, Ingres, DEC, Tandem, Sun and HP.  Microsoft joined round about 1990.

This group came up with a call level interface (CLI) that they defined for SQL databases.  It was intended to allow programming languages to access a single driver written a database vendor.  With this approach, if you had the 5 apps and 5 DBs listed above, you would need 5 drivers, and your apps would all need to have an interface provided that would talk to the drivers. 5 drivers instead of 25. 10 drivers instead of 100.

The SQL Access Group finally came up with a name for their standard. They called it Open DataBase Connectivity, or ODBC.  In addition to the CLI, this interface expected you to support some level of SQL. There was Minimum, Core and Extended. We implemented Core.

John's company was contracted to write drivers for all of Microsoft's supported ISAM databases, including comma-separated text files, DBase, Excel, and others.  This meant that they had to write a full SQL engine for these, as comma-separated text files had no SQL processing capability. John was working with a gentleman named Jim, who took on the task and did an excellent job.

In an interestingly twist, when a company called Fincentric wanted to convert their application development environment from using ISAM to SQL Server, Jim also wrote the inverse, taking ISAM commands and converting them into efficient SQL statements.

ODBC for PICK

In 1991, before ODBC even had a name, a number of us met at the India Gate restaurant on Robson street downtown, and discussed the idea of writing a SQL engine and driver for PICK. (We actually recorded notes on a napkin!)  Because of my involvement with Synex and PK Harmony, I was the communication expert.  John was the C/C++ expert with Jim as a very helpful resource on writing SQL engines for non-SQL environments, and my brother and I were the PICK experts.

We decided that my brother Antoon would take on a project working with SQL Server for Vancouver General Hospital (VGH) in order to better understand SQL, so we could determine the feasibility of what we were contemplating.

Laboratory Information Management System (LIMS)

The project was code-named LIMS. The funny thing was that Paradigm was assigned a phone number by VGH that had formerly been for some other department there.  Occasionally, the programmers would get a call from someone who didn't realize the number had been reassigned. They'd answer it with LIMS and there'd be a long pause at the other end, while the caller tried to figure out what department at VGH was called "Limbs"!

As a result of his work on this project, Antoon came up with an initial architecture and design for an ANSI 92 compliant SQL engine prototype, which he began coding in PICK BASIC.

It didn't take too long to get to where "SELECT * FROM SALES" returned a meaningful result set. That was the easy part.

ODBC vs. IDAPI - Oddball vs. Diaper

Borland created their own interface called the Integrated Database Application Programming Interface (IDAPI), which never took off beyond their own databases.  If you ran Microsoft Word's spell checker on IDAPI, you got "diaper". If you ran it against ODBC you got "oddball".

Microsoft ODBC Driver Developer Kit

We were one of the first to get a copy of Microsoft's ODBC Driver Developer Kit. In order to boost productivity, John took their API and re-wrote the header files to support C++, but expose the interfaces as C. I began writing the driver, and developed a transfer protocol that would work with Serial I/O, which was all that most PICK systems would support at that time. I used my experience with Synex, but based on a variation of the Motorola IXO transfer protocol that would work with serial I/O limitation of most PICK systems.

Microsoft had 3 levels of conformance to their API. Core, Level 1 or Level 2. We implemented Level 1 conformance.

Microsoft Releases ODBC 1.0 in 1992

Finally, in 1992, Microsoft released ODBC 1.0. Sort of... They implemented their driver manager so that it would only support their drivers for the first version.  We had been ready to release with them, but found out only weeks before they released that they had decided to implement this limitation.  I scrambled to write an Excel add-in that would allow us to use our driver to load Excel spreadsheets, so our beta customers would be able to use it.

We still managed to get a number of customers up and running with our beta version. And so our ODBC journey was launched.

We Failed the Bozo Test for ODBC

One last funny story for this post: It was an industry inside joke that if you could spell the letter of ODBC in the right order you probably knew what it was.  The unknowing would always refer to it as OBDC.

We put together a marketing brochure, and went through probably about 50 edits.  Somehow, on the very last edit, someone reversed the D and the B and we were so bug-eyed we missed it.  Our brochure went out with OBDC and we couldn't correct anyone who phoned in!

We failed our own Bozo Test!

The Changing Face of Connectivity

In the Beginning...

One of my earliest customers was First City Trust. While they had started on Microdata, they had begun to move the applications to an IBM mainframe computer.  This was an interesting exercise, as they discovered that it was faster to move data back to the Microdata and get reports off of it from there, than to ask for them on the IBM. They had an army of people (an entire floor) that were focused on "fixing" data that was broken by the application running on the IBM, and getting anything new run up took forever.

On the other hand, they had less than half a dozen programmers, including myself, who were doing these reports and keeping the applications that were still on the Microdata running.  It used to be said that regardless of price "no one ever got fired for hiring IBM".  We'd just shake our heads...

In those days, you had two options for moving data over. One was to print a report and hire someone from the local office temp pool to re-enter it. The other was to write it to a 9-track tape in a compatible format, then read it on the other system.  I did quite a bit of work interpreting COBOL programs that were used to write to the tapes, reading them, converting from EBCDIC to ASCII, parsing them, and then writing them into the appropriate places in the Microdata. Then we'd write a report and usually have it out in under a day.

Let There be Serial Data Transfer Protocols

With the advent of PC Harmony and PK Harmony, we had a new way to move data around.  For the minicomputer world, you could hook up an IBM PC through an RS232 port (also called a Serial Input Output Port (Serial I/O Port, or just Serial Port for short).  These ports initially would allow you to push data through a modem, or directly cabled to a terminal or PC.  Directly connected, you could connect at 9600 baud. This equated in really rough numbers to 8 bits per 10 to 12 baud, so you were getting under 1K per second.  Over a modem, you'd be lucky to get 1200 baud.  A good typist was faster than a modem.

Over time, the quality of the serial ports, terminals and modems all improved, until you could finally get 56K baud modems and 19,200 baud terminals.

The IBM PC came with a serial port interrupt handler in the BIOS that could barely handle 1200 baud on a good day.  In order to get any more speed, terminal emulator writers had to override and replace that interrupt handler with their own that was optimized.

Serial Ports had another problem. They were unreliable. Even when you had hooked the interrupt handler, there were times when enough other interrupts took over the machine and you would drop characters. Or electrical interference, or line quality for a modem would result in corrupted data.  There were parity bits in the low-level protocol that would attempt to help you determine if corruption had happened, but they were typically just even or odd parity. If your corruption flipped two bits, the parity bit would look right, but the character would be bad.

So, in order to ensure that data got through, we developed a proprietary error correcting data transfer protocol.  We checksummed the whole data package, put a length prefix on it, and then checked the checksum at the other end.  We also ensured that every character was echoed before we sent the next one, and would time out and retry.

This was a lot of work, for a relatively slow process, but it beat having someone re-key the data, so it was used by many Business BASIC and Pick BASIC users for many years.

Then There was Netware

For a long time, Novell's Netware was king in the networking world.  I remember when they release a new feature. You could send data over the Netware network, and it would connect to a Serial Port concentrator, that was hooked up to your minicomputer.

That last bit of connectivity typically had the same 9600 baud or maybe 19,200 baud limitation, so it wasn't blindingly fast, and it also had the reliability issues that the minicomputer brought to the table, so you still needed the error correcting protocol, but it had one significant benefit.

Prior to this, if you had a factory, and you needed to put a new terminal in it, you had to run a Serial Cable all the way from your minicomputer to where the terminal needed to be.  It didn't matter how many cables you had already run all that way, you still had to do it.  And Serial I/O was only certified for a certain distance, so unless you paid for expensive repeaters, you were not going to have very good reliability.

Now, you could run a network out to the factory floor, and by dropping a line off the network to a new PC running terminal emulator software, you could add it with very little cabling and very little impact on reliability.  In order to use this network stuff, however, you had to implement a protocol called Netware Asynchronous Communication Services (NACS).

We had a few customers who used this interface, but networking was still fairly new, and required deep pockets to implement. Most Pick users didn't have those deep pockets.

An Explosion of Networks

With Novell initially leading the pack, networking continued to advance. When implementing networks in the early days, you would look at ARCNET, Token Ring, Ethernet, and then you would have to consider topologies. You could choose between Star, Bus, Ring or Tree topology.  While all this was going on, the early internet was being developed.  It used Ethernet, and used a Mesh topology.

Nowadays, the internet uses a combination of Ethernet, Wireless (WiFi), and cellular data technologies. The latter two are gradually outstripping the connected Ethernet for end-user devices.

In the mid-1990s at Liberty, we hired a student called Pardeep, to assist us with support. In speaking with him, he suggested we consider using Winsock libraries to write a TCP/IP socket driver.  Many of the minicomputers we were working with supported TCP/IP based Telnet services, and since he was familiar with the Berkeley Sockets, which Winsock was based on, we asked him to write us a socket driver. Our LWinSock.dll dynamic link library is used to this date at a number of companies including two fortune 500 companies, in our ODBC driver.  We also ported the same driver to JDBC, and used a variation of the code in our OLEDB driver and managed provider.  That little socket driver has seen a lot of use over the years! The beauty of this was that we were able to completely drop all the error correcting protocol and simply acknowledge receipt back to the sender.  We could also send large blocks and it would generally be handled provided we didn't overrun the buffers at the target end.

Was That a Bagle or a Nagle?

One thing that we noticed after a while with our socket driver, was that it would perform very slowly when smaller packets were being sent to the server from the PC. I remember doing careful analysis and concluding that sometime between us requesting a send and the server receiving it, something was introducing a precisely 200 millisecond delay.  It took a lot of digging to unearth the culprit, but we finally did.

When they first introduced Telnet protocol, the implementers discovered that if you had a lot of Telnet users typing characters, by computer terms, there was a large gap between each character, so it would send a character at a time in a packet. The overhead in a packet of 1 byte was huge and wasteful, so a gentlemen with the last name of "Nagle" came up with a strategy. Simply put, as the human typed, the client end of the Telnet program would gather up typed characters and not send them until either a certain number of characters had been gathered, or 200 milliseconds had elapsed since the first character was typed. This was a bit like holding the elevator door and poking your head out to see if anyone was coming before letting the elevator go.

This was called the Nagle algorithm and turned out to be the culprit.

It seems they assumed, if you were opening up a socket to a Telnet server that you would want the Nagle algorithm turned on. We turned it off and the small packets flew fast and furious!  Lesson learned.

Telus PureFibre ™

Today I have Telus PureFibre, and get well over 150 megabits per second. That's roughly 150,000,000 baud!  I can connect wirelessly or with a physical cable, and generally, the other end is the limiting factor for speed.  Reliable data transfer is pretty much a given. If your wireless connection goes down, you will get an error and may have to restart your transfer, but in general, if your network stays up, the data will get there, reliably and fast.

Hotels often provide internet connectivity, but then have hundreds of guests sharing the same point of presence, and it's generally fat enough for a family of 5 with 2 watching Netflix. It doesn't handle hundreds of users at once.  That said, if you can find a decent Starbucks or other point of presence, you can generally get good, fast, connectivity today.

I know that new protocols and new technologies are in the works.  I can't wait to see what the future holds!

Microsoft's Office Productivity Software

An Unexpected Surprise

Some time after we released PK Harmony at Synex, several of the people I had known there left, and went to work for Microsoft.

One day, I got an unexpected package in the mail from Microsoft. It was a pretty big box.  In it was software diskettes and user guides for two products. One was Microsoft Excel version 3.0 and the other was Microsoft Word 1.2.

They both came with hard cover user guides that were about the size of a standard letter sized paper sheet.  They had complete descriptions of all the commands and options in the books.

Vastly Improved Word Processing

Prior to getting these programs, I had Lotus 1-2-3 for spreadsheet work, and when I needed to create documentation, I used a primitive program called RUNOFF which had been ported to Microdata.

At Synex, we had people doing technical writing using products like Pagemaker and Ventura Publisher, so I was familiar with WYSIWYG editors, but they were expensive and I didn't have one myself. That all changed when I got MS Word.  I got some great pointers from our technical writers on how to leverage styles, which I use to this day.

In high school I had taken a typing course, largely to fill in a gap, but because I thought it would be cool. The result was that I was one of the few programmers who actually knew how to type (most did chop-sticks typing - two index fingers) and as a result I could type very quickly.

Microsoft Office Suite

In a few short years, Microsoft bundled Word, Excel, PowerPoint and for the professional version, Access, all in a single bundle called Microsoft Office.  Microsoft continued to send me free versions to work with, including a standalone Word 6.0 version that could install on DOS or OS/2 Presentation Manager. (You didn't actually need Windows to run early versions of MS Word.)

Because of PK Harmony, and the need to support Mail-merge and other automation features, I began to become quite expert in these desktop productivity products, like Lotus 1-2-3, WordPerfect and others.  We did a lot of macro programming to automate loading data into these programs from people's Business BASIC and PICK systems.

I could write technical documentation and get it 75% of the way there. The technical writers did the last 25% of the formatting and cleanup, giving it that professional shine.

Object Linking and Embedding and Component Object Model

As Microsoft developed their operating system platform, a new approach to integrating applications began to appear.  Through Object Linking and Embedding (OLE) and Component Object Model (COM) an application could expose its functionality so that another program could easily integrate with it.

Just a few years ago, I used the Microsoft Excel COM automation libraries to write a C# program to extract insurance rate data from an Excel spreadsheet that the executive had used to model their next period's rates, and push it into their multivalued mvBase database that ran their insurance line of business. The program I wrote ran through and validated that the spreadsheet had the right tabs, that the tabs had the right columns, and that the columns had the right data and that the data relationships made sense.  If anything was wrong we reported a highly descriptive and detailed error and stopped. If all was good, we used our own libraries, licensed by the user, to construct multivalued records and write them to the appropriate staging file (all done in C# - we had a library of multivalue dynamic array handling methods written in C#). The customer would check that file and validate it before replacing the live file with the data we uploaded.

Before OLE and COM, we would have required access to the source code for Excel, or would have had to export the data using a macro, or most likely would have hired someone to data enter it all.

Microsoft Office Today

Today, we have Office 365. While you get access to fat-client programs that you install and run on your device, you can store your data in the cloud, and use slightly limited web versions of the software from anywhere.  What's more, my subscription allows me to have Word and Excel running on my laptop, my Mac, my iPad and my iPhone, and they can access data on OneDrive, so they can all access the same information. I can switch where I'm working and continue where I left off!

As you may have guessed, I'm a great fan of the Microsoft Office suite of products.

Microsoft MVP Reconnect

A New Community from Microsoft

The other day, my colleague, who is a current Microsoft MVP, told me about a program that Microsoft was putting together to bring past MVPs together in a community. The program is called MVP Reconnect.

I was awarded Microsoft MVP status for two different areas: The first area was ODBC, the second was for Win32.  Both of these are far enough back that Microsoft doesn't have records, so I had to pull up and email them scans of the various letters I had received in 1995 and in 2000 for my MVP awards.  About a week ago (May 2018) I was accepted in the MVP Reconnect community.

® Microsoft is a registered trademark of Microsoft Corporation in the United States and/or other countries.

MVP for ODBC Contributions

Back in 1995 when the Microsoft ODBC forum was hosted on the Compuserve network, I found that I answered more questions than I asked when I was on there. As a person who developed and deployed ODBC based solutions, and write ODBC drivers and back-end server software to serve the SQL syntax that was key to ODBC, I found that I had a unique understanding of how the middleware worked, so I was able to help many people with answers, regardless what driver or applications they were using.

Back in the day, some of the most prolific members of that forum included Microsoft employees: Murali Venkatrao, and Mike Pizzo, and non-Microsoft people: Lee Fesperman, Ronald Laeremans, Charles, McDevitt, Rob Macdonald, Mark Edwards and Dale Hunscher.  And many more whose names don't come to me now.  Between us we helped a lot of people in the early days of ODBC.

Microsoft release ODBC 1.0 in 1992. Unfortunately, the first ODBC-compatible versions of their Office suite really didn't play well with anyone else's ODBC drivers. Their applications only supported their own drivers. It was the next version of Office that finally provided real ODBC connectivity.

By 1995, my contributions were recognized, and I became an MVP (the program started in 1993, I believe.)  This award continued until they moved all Microsoft forums from Compuserve to MSN in 1998.  Unfortunately, I was running on Windows NT, and you needed Windows 95 to access MSN, which had a proprietary front-end application, so I got knocked off the network. Even if I had had access, no one posted any questions to the MSN network.  I continued to respond on the Compuserve network and the ODBC forum owners gave me free access for several years beyond that because of my contributions.

MVP for Win32

Then in 2000, I was awarded MVP status again, because of contributions I was making to Win32 developer forums.  That was before .NET was released, and I was doing a lot of C++ development.

MVP Reconnect

With all the attendant perks, and the chance to reconnect with some community minded industry experts, this was a pleasant surprise! I'm looking forward to interacting with this community!

Playing Games on Computers

My First Computer Game

My first computer game really wasn't my own, it was my brother's.  I had two brothers, 6 and 7 years older than me, and the younger of the two had a Tandy Radio Shack TRS80 computer. It used our TV as a monitor, had a built-in keyboard, a game controller, and a cassette interface to back up and restore data, and to play games from.

It had a space invaders game where you went around the galaxy shooting up alien ships, periodically returning to your space station to refuel.

When I finally got a chance to play it, at about 14 or 15 years old, I had a nefarious plan.  I shot up all the aliens except one, figured out where he was, then went back to the space station and refueled. As I left the space station I blew it up.  A message about a pending court martial scrolled across my screen, but it let me continue on. 

So I continued on to where the last alien ship was and blew it up.  My plan had worked. I controlled the galaxy as supreme ruler (basically the only one with a space ship with weapons!)

Character Based Games

When I finally got into computers myself, it was initially on mini-computers with ASCII terminals.  Many of the terminals when I first started could not handle lower-case characters, and don't even think about graphics, so you were limited in what you could do.

I remember one game we played on our Reality systems, I think it was called Hammurabi.  You were buying and selling on the open seas in the ancient world. There were money lenders, including Hammurabi, who'd charge ridiculous interest rates and would chase you down looking for payment periodically.  

My brother-in-law got to playing the game, and discovered that when the money-lenders offered to lend money, you could type a negative number. They would basically have borrowed from you, and continued to charge (themselves) usurious interest rates and the money in your account would grow dramatically, with compound interest!.  He got so rich, his wealth finally exceeding the numeric capability of the computer and it crashed the game!

Microsoft Flight Simulator

When I finally got a personal computer, one of the first games I got was Microsoft Flight Simulator. This was an amazing game.  

Apparently real pilots liked to use it, because it was so realistic in so many ways.  You could select the type of plane you wanted to fly, and my pilot friend explained that the game would simulate both the correct cockpit controls and other quirky characteristics of the plane.

I loved that game and enjoyed quite a few hours on it. I was extremely disappointed when I finally upgraded from my old IBM PC XT with it's 4.77 megahertz clock-speed to an IBM AT clone with a 16 megahertz clock speed, and discovered that the game based its timing, not on the clock, but on CPU instruction times.  At about 4 times faster, the game would get you up in the air and tunneled back into the ground so quickly you couldn't really control it!

Games with my Son

I wasn't really into games that much, but as my son got a bit older, I began to look for games we could play on the computer and enjoy some family time. It was the latter half of the 1990s and my computer had progressed and it was now capable of VGA graphics and Soundblaster sound, so we began to find some really fun shareware games.  While many of these would support game controllers, finding game controllers that were any good for a PC was a pursuit we gave up on. 

We had several Commander Keen games, including our favourite, Keen Dreams. Little commander keen is wandering about in his dreams in PJs and slippers and defeating slugs to get back to his bedroom.

Another was Wacky Wheels. We could actually both play it at the same time by taking different sides of the keyboard.  You were driving racing cars as hedge-hogs and shooting mini-hedge-hogs at the other driver.  It was a blast!

We also found a number of Duke Nukem versions, and really enjoyed them.

Similar to Flight Simulator, as PCs progressed, these games "broke", yet using the virtualization of the PC, some very smart people came up with a program called DOSBOX that mimics the behaviour of an old DOS based PC, with Soundblaster sound cards and VGA graphics.  I can still play the Commander Keen, Duke Nukem and Whacky Wheels games using DOSBOX, and occasionally do (my son does, too!)

A friend gave us a copy of a game called Arthur's Teacher Trouble, which was another cute game.  We discovered some hidden Easter eggs in that game and it was a great favourite for our entire family, including nieces and nephews, for a long time, until Windows upgrades broke it, too.

As a result of my second stint as a Microsoft MVP, I got a copy of Monster Truck Madness, and more fun with my son and nephews ensued. There was a level where, if you went off-road, you could crash into cows and push them around. They'd "moo" and in the background a farmer's voice could be heard saying things like "Hey! You kids! Get off my farm!"  Another level had a trailer park in the desert with outhouses. If you crashed into an outhouse, you'd hear people inside yelling at you to stop.  As you can imagine, my preteen son and nephews got some great giggles from this.

Riddle of Master Lu

One of the people it was my pleasure to work with was my friend Pardeep.  After working for us for a time, he went on to a company called Sanctuary Woods, and helped them create a game called "The Riddle of Master Lu". This was a first person player game where you were Robert Ripley, looking for some treasure.  He was very handsome and got chosen to play the part of a monk in the game.  It was really cool to have a game where one of my former colleagues was an actor and a developer!

My Favourite Computer Game of All Time

And so we come to my favourite game of all.  About 25 years ago, I discovered a game at a trade-show. The graphics were amazing and it looked like a really fun first-person exploration game. It was called Myst.  

My son and I both checked it out, and quickly realized that our current PC simply didn't have a powerful enough graphics card. Worse yet, our hardware was incapable of supporting any of the required graphics cards.  Undaunted, for the first time ever, I upgraded a PC for personal reasons, not business reasons.  I got a new PC that would support a required graphics card and had them install the card for me.  Then I bought Myst. And Uru.  And all the follow-up games that came out from Cyan in that series.  We loved it!

The graphics were stunning. The story-line was amazing. The game was challenging, interesting, and engaging, so that you felt it was really you in there.  The realism of the story line and how things progressed was very good. And as a parent I was quite pleased that it was pretty-well completely non-violent.

As operating systems matured, and the user interface layer was taken out of protected mode (to mitigate the impact of exploits to the graphics layer), the games were essentially broken. I understand that they were written to work with a limited set of specialized graphics cards. It became increasingly difficult to find computers that would run the games, and they fell out of the mainstream.  There was one attempt to do an online version of the games, that I don't think went very well.

The good news it that Cyan is now running a Kickstarter campaign to create a 25th Anniversary Myst collection.  Needless to say, I'm backing it and looking forward to getting my copy of the series this summer!