SQL Compact Command Parameter BLOBs

The problem du jour is to make the CCommand class accept BLOB parameters, something that I have been thinking about the last week or so. Let me recap what's at stake.

When executing a SQL command with parameters, the OLE DB standard allows only for a single accessor handle to be provided by the consumer. This accessor handle is created to bind the command parameter data to a data buffer, just like in a normal rowset. By just having one accessor to handle bind the whole set of parameters means that we cannot bind more than one BLOB with SQL Compact, due to its OLE DB provider limitations. In order to have a general solution, we must bind the BLOB parameter as regular columns but we must know how much memory to allocate for each BLOB parameter, and for each instance of the command execution.

After some thought I concluded that there are two acceptable approaches:

1. Ask the user to provide maximum acceptable size for each BLOB. This is what the ADO .NET provider does and essentially delegates this responsibility to the user. Do you know in advance what is the maximum size of each BLOB that you are going to feed as a parameter? If so this is acceptable, if not... The great advantage of this approach is that you can make an apriori parameter binding (and make it only once).
   
2. Be smart about it and recreate the binding when needed. This requires keeping track of the current BLOB binding sizes and increasing the binding array when needed. This approach frees the user from having to specify a maximum size for the BLOB columns but requires that the binding be rebuilt when a larger BLOB is used.

Because the second option seems to be hardest to implement, that is exactly the one I chose. See you in the next post with the code!

SQL Compact Command Parameters - II

After a bit of work, here's the new version of the OLE DB Client library with basic support for command parameters. The sample code implements a very simple query:

SELECT * FROM Customers WHERE Country = @c

and executes it by setting the parameter value to 'Germany' (using the same sample database).

The first change you will see in the sample code is that I downgraded the solution to VS 2005 format due to some requests to do so. VS 2008 users can easily upgrade the solution, so this is a minor issue. For convenience, I'm also packaging everything in a single ZIP file. Also, all the samples from now on will deploy to the same device directory (typically in \Program Files\OleDbClient).

Executing a SQL Compact command with parameters is quite simple. In a nutshell:

command.SetText(pszCommand);
hr = command.Prepare();
if(FAILED(hr)) ...
command.SetParam(1, _T("Germany"));
hr = command.Execute(rowset);

Remember from my last post that parameterized commands must be prepared before executing. If you look under the hood, you will see that parameters are bound (using code very similar to CRowset's column binding code) in the Prepare method. You can only set command parameter values after the Prepare method executes successfully.

This code still has some limitations, and the most notable one is the inability to handle BLOB parameters. When executing SQL commands, the OLE DB provider only accepts one accessor handle to bind all parameters. As we have seen in previous posts, this prevents us from binding more than one BLOB parameter. The only solution I devised for this when using the ATL OLE DB Consumer Templates was to force the BLOB parameters to have a maximum lenght and bind them as regular columns. Of course, this implies a change to the parameter binding code and some user intervention...

Sample code: OleDbClientLib.zip

SQL Compact Command Parameters

Command parameters can be thought of as value replacements or variables for SQL commands. You use a command parameter when you need to execute a SQL command multiple times changing one or more values on each execution. Instead of rebuilding the command text every time the command is executed, your code merely changes a variable value. This means that your code changes from this:

SELECT * FROM Customers WHERE Country = 'Germany'

to this:

SELECT * FROM Customers WHERE Country = @p

Note that the command parameter (@p) replaces a literal value that will change every time you execute the command. Command parameters have some interesting advantages over string concatenation:

• Command parameters are strongly typed;
• Command parameters make the command run faster;
• Command parameters are safe because they avoid SQL injection attacks.

Suppose that you need to pass a date value as a parameter to a SQL command. How exactly do you format a date value in order to be correctly interpreted by the query processor? When using command parameters you just don't are about this - you just feed the parameter with the binary value and it atakes care of the rest, no string formatting contortions needed.

The OLE DB provider requires that parameterized commands be prepared (compiled) before executing. Contrary to the managed ADO .NET provider, the OLE DB provider does check the command for errors and enumerates all the given parameters and even infers their data types (your code can then read this informtion and use it). Preparing a command with the ADO .NET provider for CF merely marks the command as requiring preparation before executing. Either way, a prepared command runs faster than an unprepared one.

Finally, with command parameters there is no way to let your user change the syntax of the underlying SQL command. If you use string concatenation to build command strings (with user input), you risk exposing your code to SQL injection attacks that may break your application code or even its security.

Since version 3.0, SQL Compact supports named parameters meaning that you can give your command parameters some meaningful names. All command parameter identifiers must be prepended by a '@' character. If you are still using SQL CE 2.0 then you are out of luck as all command parameters are identified by a single '?' character and distinguishable by their ordinal position only.

After this brief introduction, I will look at the details of implementing a SQL command parameter management code and some of the finer details of binding, such as what to do with BLOBs.

Very fast row counting on SQL Compact

What's the fastest way to retrieve a table's row count in SQL Compact 3.0 and 3.5? Use the sp_show_statistics stored procedure! Pass your table name as the first parameter (the second is optional) and you will get back a few statistics about the table and its indexes, so more than one row might be returned. The row count is on the ROWS column (ordinal 3), an unsigned 32-bit integer.

Note that this is not supported in versions 2.0 and earlier.

Meet me at Tech Ed EMEA 2008

Meet me in Tech Ed EMEA 2008 where I will deliver a presentation about "Microsoft SQL Server 2005 Compact Edition Ultimate Performance Tuning" (Windows Mobile track). See you there!

New column binding and commands

As promised in a previous post, I changed the way columns are bound in the CRowset class so that the class retains all the important column information retrieved from the IColumnsInfo interface, but also the binding offsets. You can see all of this at work in the new version of CRowset in the sample files.

You will also find the first implementation of the CCommand class, an abstraction over the OLE DB Command object. This first implementation is very simple and allows only for execution of a command text (no parameters yet) returning a CRowset and/or an affected row count.

Using a CCommand object is very simple (much like using CTable):

CCommand command(m_session);

command.SetText(_T("SELECT * FROM Employees"));
hr = command.Execute(m_rowset);

Start by declaring a CCommand object passing the open CSession object as parameter. Next, set the command text and then call execute passing a CRowset object as parameter. You can now use the rowset to navigate the rows and retrieve data.

Note that the returning rowset may not have a bookmark (not in this case, anyway) so you must be careful how you index the resulting columns. Remember that CRowset::GetColumnCount returns the total number of columns including the bookmark, if present. Also, columns are referred to by their ordinal position and the first data column is always at ordinal 1, so if you are iterating through all columns using GetColumnCount, make sure you always start at 1 and also add 1 if the rowset does not have bookmarks (see CRowset::HasBookmark).

Finally, if you don't set any specific command properties (as in this case) you get a forward-only cursor, so there is no going back...

Sample files: EditRow5.zip, oledbcli4.zip

P.S.: Don't try to use the EditRow sample to edit rows coming from the command rowset! I just adapted the sample to show how easy it is to display command data. The resulting is read-only.

OLE DB Commands and SQL Compact

OLE DB Commands are created through the IDBCreateCommand Session object interface. Just find your Session object and call QueryInterface using the IID_IDBCreateCommand GUID. This interface just exposes one method - CreateCommand - that you use to create OLE DB Command objects. As with most COM and OLE DB objects, several interfaces are exposable for a given object subject to provider implementation. For SQL Compact the following interfaces are exposed:

• ICommand - Execute the command through this interface. As I noted in a previous post, the Cancel method is not implemented in SQL Compact.
  
• ICommandPrepare - Use this interface to Prepare or Unprepare the command. Command preparation is useful when executing the same command more than once and is required when using command parameters. Contrary to what happens with the ADO .NET provider, the Prepare method actively validates the SQL Command against the existing database, returning errors for incorrect syntax.
  
• ICommandProperties - Gets or sets the properties associated with the Command.
  
• ICommandText - Gets or sets the command text. This interface inherits from ICommand.
  
• ICommandWithParameters - Manages command parameters (we will have to implement a data binding scheme similar to what exists in CRowset). SQL Compact allows only for input parameters, not output.
  

As you can see, this is enough to run SQL commands against SQL Compact's Query Processor. For now, we leave this here as a reference while I go back to work and start implementing these features in the OLE DB Client library.

BLOB update code

The sample code I'm posting today illustrates the techniques I described on my last post. Please note that this is really a sample - I will change it in the next incarnation in order to be more effective when binding data. Right now you can see a bit of a mess when binding columns because both the DBCOLUMNINFO and DBBINDING arrays are being used to find the column information addresses on the data buffer. Next version will have a unique array with all the required info to correctly and accurately describe each column. Enough of this and let's look at the code.

First of all, take a look at CRowset::SetValue. Here you see how text strings (only!) are written and how the code decides if the column is a BLOB or not. If we are writing to a BLOB column, the code creates a CBlobStream object and writes the string contents to it. Upon success, a pointer to the CBlobStream object is stored in the rowset data buffer. When this data is written to the provider (by calling IRowsetChange::SetData on the BLOB column's accessor handle), the provider calls the CBlobStream::Read method to read all the string contents into the BLOB storage. When it is done with it, the provider calls the Release method and the CBlobStream deletes itself - neat.

Finally, take a look at the CRowset::Insert, CRowset::Delete and CRowset::SetData methods. These implement the changes I described before. Also, note that these methods will work even if the IRowsetUpdate interface is not exposed.

Next up - After rewriting most of the column binding code, I will start talking about commands.

Sample files: EditRow4.zip, oledbcli3.zip

Updating rows with BLOBs

As I have previously discussed, we need to make some substantial changes to our code if we want to update rows with BLOB columns. So how exactly do we update these things? The first thing we need is to implement an object that behaves like an ISequentialStream COM object. Our code will create one such object per BLOB column, store the data in this object and then store a pointer to this object in the rowset buffer. When updating data, the OLE DB provider will read it (just like our code reads the storage objects coming from the provider) and then releases it. To make our life easier, our implementation of this object should delete itself when the reference count reaches zero. This way we will not have to worry about deleting it later.

After setting all the row data (BLOB and otherwise), we can insert or update it. If you want to delete a row it is pointless to write data to it. If you remember, I said that when using BLOBs with SQL Compact we must use the deferred update mode by requesting the IRowsetUpdate interface. Here's how the code changes for each operation (insert, update or delete):

Insert

1. Call IRowsetChange::InsertRow - this sets the non-BLOB data;
2. Call IRowsetChange::SetData for each BLOB accessor;
3. Call IRowsetUpdate::Update to commit the changes.
   

Update

1. Call IRowsetChange::SetData for each accessor.
2. Call IRowsetUpdate::Update to commit the row changes.
   

Delete

1. Call IRowsetChange::DeleteRow.
2. Call IRowsetUpdate::Update.

Note that if any of these processes fails before calling Update, you should call Undo to revert any changes made to the row. I'll show you the code in the next post.

ICommand::Cancel is a dummy

No, I'm not going to talk about the ICommand interface right now, but I just found out that the ICommand::Cancel implementation on the SQL Compact OLE DB provider is a dummy. Good to know!

Reading BLOBs

My first approach to the BLOB-handling code took a bit longer than expected due to my aging brain. Please understand that the code I'm publishing today is incomplete and will change in the future. I hope that it helps you in understanding what it takes to read BLOB data through the SQL Compact OLE DB provider.

There are two important changes to the code since last post:

1. Columns are indexed by their ordinal.
   
2. Columns are bound using a two-pass approach - On the first pass the non-BLOB columns are bound using the same accessor handle. The second pass binds the BLOB columns allocating an accessor handle for each column.

The bulk of the code that reads the BLOB into memory is on the new CRowset::GetValue method. Note how the column accessor handle is used to retrieve the ISequentialStream pointer and how it is read into a CString. You have to consume all the data immediately because you cannot keep that stream object alive very long, especially if you want to read or write to another BLOB.

Most of the complexity you see on the code (the difference between binding and column indexes) stems from the fact that columns are being moved around in order to make binding simpler. When life was simpler and there were no BLOBs, there was a direct relation between column and binding entries and there was no need to map between the two. Now that we are moving binding entries around, we must make sure that a mapping exists between both arrays (see the GetBindingIndex method).

Next post will handle writing data to the BLOB.

Sample files: EditRow3.zip, oledbcli2.zip

CStdDialogImpl and the OK button

Are using CStdDialogImpl as the base class for your WTL dialogs? Are you having difficulty hiding the "ok" button? Here's a simple solution:

Derive your dialog class (say CMyDialog) from

CStdDialogImpl<CMyDialog, SHIDIF_SIZEDLGFULLSCREEN | SHIDIF_SIPDOWN>

On the OnInitDialog handler, call:

SHDoneButton(m_hWnd, SHDB_HIDE);
ModifyStyle(0, WS_NONAVDONEBUTTON, SWP_NOSIZE);

Make sure you change all the calls to CStdDialogImpl<CMyDialog> methods with the new base (you can use a typedef for that).

Let's talk about BLOBs

So far we have seen how to edit table data using the IRowsetChange interface. The provided sample managed to edit a simple table where all columns had limited size. Table edits are performed in a single step, where all the data is read or written at once, and all the data items are sequentially stored on the same data buffer. This scheme works very well but does not support BLOBs.

A BLOB is a different beast because it can store a very large amount of data. In SQL Compact there are only two types of BLOBs: the image (variable length byte array) and the ntext types (variable length character array). These can be really large and the maximum theoretical size of a BLOB field is 4 GB in SQL Compact (the same size limit for the whole database).Contrary to smaller data types, you may not be able to store the whole BLOB data in the device memory, so we need a different approach to handling these types.

The OLE DB standard allows two different ways to bind BLOB data to your accessor: by reference or by value. SQL Compact does not allow by reference binding, so we are limited to binding these by value? What are the differences between both types of data binding?

• When binding by reference, the OLE DB provider allocates memory for each bound BLOB field and fills it with data. A pointer to this data is stored in the accessor buffer along with its status and size. Column data is accessed through an extra pointer indirection and our code must make sure that the BLOB data buffer is released when the data is not needed anymore.
• When binding by value, we can either specify (force) a fixed column size or bind through a storage interface. The first approach allows your BLOB data to be manipulated just like any other column but forces you to put a limit on how much data you will handle per BLOB column and this defeats the whole purpose of using them. Alternatively (and most commonly) your code will request that the provider creates a storage object to handle the BLOB data (when reading - when writing data your code must do this). When data is read ftom the table, the provider stores a pointer to either an ISequentialStream or an ILockBytes interface and then your code can read the data through these. When writing, your code must create one of these objects per BLOB column.

The last approach - binding by value through storage objects - is the recommended approach, but it becomes harder to implement when your code is handling more than one BLOB per row. In fact, SQL Compact will only manage (read from or write to) one storage object at a time, and if you ask it to instantiate more than one you will get an error. So how do you read from a table (or query) that returns more than one BLOB column? And how do you write the data? Well, you do so one at a time and this means that your code cannot read nor write the whole data in one single step.

The first change we must make to the rowset code is to add more accessor handles. The first accessor handle will bind the non-BLOB columns and these can be read and written to just like we are doing right now. For each BLOB column the rowset must allocate an additional accessor handle and make a note of it.

When data is read, the non-BLOB columns are all read, but none of the BLOB columns are. When your client code requests data from a BLOB column, the rowset must know that this is a special case, retrieve the colsumn's accessor handle and call IRowset::GetData on that handle. On the data buffer, the OLE DB provider will write a pointer to one of the above-mentioned storage objects (it is chosen by the rowset code when the data binding is created) and your client code must read through it and release it when no longer needed.

When you need to write data back to the table, your client code will have to create one storage object per BLOB column, store the respective pointers in the data buffer and engage in a multi-step update (new row insertion or existing row update). This means that your code will store the non-BLOB columns first and then store the BLOB columns one at a time, but to be able to achieve this the rowset must be in delayed update mode. You specify this mode by setting the DBPROP_IRowsetUpdate property to VARIANT_TRUE and the rowset now exposes the IRowsetUpdate interface. When this interface is exposed, you can call IRowsetChange::SetData on the allocated accessor handles without having to worry about having the row in an inconsisten state (due to table constraints, for instace) while you are updating data. When your code is done setting data through all the accessors, just call IRowsetUpdate::Update to update all the changes at once.

This will be easier to understand through a sample, which I will publish on my next post.

Using the OLE DB Client code

I was recently asked for permission to use the OLE DB Client library code in a third party application. One of the blog readers asked me about a specific license to go with the code and I pointed him to the CPOL, the permissive license used by most Code Project articles. I'm actually planning to convert the material from these posts into a series of articles to be published in Code Project under this license. This is a very simple license: you can use the code for whatever purposes you like (even in a commercial application) and the only thing you are required to do is to acknowledge who wrote the original code (see the CPOL details).

Now that you know this, should you use this code? If you want to use it to learn about OLE DB and the SQL Compact native provider, please go right ahead and use it (and change it, and tweak it and experiment with it). As to using it in a commercial application, I would say "not right now" for two reasons:

1. The code has not been extensively tested. Although I plan to publish a bug-free (ah ah ah!) library, I cannot take resposibility for any use you make of it. You will be on your own.
   
2. This code will change. Don't start writing your own code based on this library expecting that the library interfaces will not change. They will, and your code will fail to compile.
   

Now that this is clarified, let's have a look at BLOBs and what we need to support them.

Moving to a DLL

Can I put the OLE DB Client library in a DLL? I tried but got some compiler warnings (C4251) concerning some of the ATL class templates (CComPtr and CString). The reason is that other classes might inherit from these and template definitions are not always exportable.

To solve this problem, I replaced all references of CComPtr in the protected interface with raw COM pointers, which required me to implement some missing destructors. I also removed the protected CString in the CTable class and now the code cleanly compiles into a DLL.

You can download the rewritten sample code (EditRow2.zip and oledbcli.zip) and test them. Make sure that you place the OleDbClient project in a directory with the same name and at the same level of EditRow.

Updating data

There are two ways to update data using OLE DB: using SQL DML commands (which I will look at in later posts), or through the IRowsetChange interface exposed through a base table IRowset. It supports the basic methods to insert new rows (InsertRow), update existing rows (SetData) and to delete rows (DeleteRows) and is only exposed when the consumer sets the DBPROP_IRowsetChange property to VARIANT_TRUE.

To illustrate this type of data update, I put up a sample that allows you to edit the "Suppliers" table from the Northwind sample database. The list view display code has been changed in order to support a load-on-demand cache of rows, stored in their native (CRowset's) format. The cache items are preloaded when the list view notifies the parent through the LVN_ODCACHEHINT notification message. Row bookmarks are stored along the row data, so we can now very quickly position the rowset cursor in any row so we can edit its data.

Note that when modifying any data, we have to make sure that both the bookmark and the row ID columns are not touched because both are managed by the engine (the row ID is an IDENTITY column that gets automatically filled in upon insert). Apart from this, setting data is just a matter of copying the new row value to the rowset and correctly setting the column status and length (optional for fixed-length types).

This code does not support BLOBs yet. I will look at these in my next post.

Sample project: EditRow.zip (1.29 MB)

Bookmarks

So what is a bookmark and why should you care? To put it simply, a bookmark is a row unique identifier on an open rowset that allows for very fast random seeks. In order to get a rowset with bookmarks, you must set the DBPROP_BOOKMARKS property to VARIANT_TRUE. The resulting rowset has the bookmark value at column ordinal zero and its size and type varies according to the database engine. On all versions of SQL Compact the bookmark type is a 32-bit integer (DBTYPE_I4 on SQL CE 2.0 and DBTYPE_UI4 on 3.0 and 3.5).

There are two important bookmark restrictions that you must be aware of:

1. The bookmark value is valid only for the particular rowset that generated it. You should not store this value for later use after the rowset closes because the bookmark values will be rendered invalid. Even if you open the same rowset again, there is no guarantee that the bookmark values will be the same. (Ok, now that I wrote this disclaimer I can tell you that apparently on base table cursors, SQL Compact does seem to reuse the same bookmark values. Nevertheless don't reuse them, please.)
2. There is no way to know a bookmark's value before loading the row data to memory. This means that if you are using bookmarks to navigate in your rowset, your code must visit the row before knowing its bookmark value, just like in a book.

So why should you care about bookmarks? I have found a very interesting use for them: editing table data without using unique indexes or primary keys. The bookmark gives you an exact and unique identifier to any row and the rowset pointer moves very quickly to that particular row.

Before you can seek to a known bookmark, your rowset must be created with the DBPROP_IRowsetBookmark property set to VARIANT_TRUE. Doing so exposes the optional IRowsetBookmark interface on the IRowset, and you can use it to very quickly position your rowset pointer to any valid bookmark by calling the PositionOnBookmark method.

Before looking at BLOBs, the next post will show how to edit data using a base table cursor and bookmarks.

Sample Databases

Isn't this nice? I have been publishing the "OpenTable" sample code but forgot to publish the sample databases. My apologies! You can find the three databases (test20.sdf, test30.sdf and test35.sdf) on this zip file: testxx.zip

Please note that the sample has hard-coded the location of the SDF file in the device root. You can change this, of course.

Database Schema Rowsets

Let's make a small detour here, now that we have mastered the IRowset, and take a brief look at database schema rowsets. The OLE DB specification requires that providers expose a set of rowsets that contain information about the database schema. You can find information about the existing tables, columns, constraints and more (depending on your provider).

Rowset schemas are created by IDBSchemaRowset, an optional Session interface that is supported by SQL Compact, through the GetRowset method. When calling this method, you must specify the rowset schema GUID and an array of VARIANTs containing restrictions that will filter the output of the rowset (each rowset has its own set of restrictions). Each restriction value corresponds to a column in the schema rowset and is either empty (VT_EMPTY), meaning no filtering will occur, or has a value of a type compatible with the column's (see here) by which the rowset will be filtered.

To illustrate this, I adapted the last sample code so that it displays the TABLES schema rowset. The very simple implementation is on the CSchemaRowset (base class) and the CTablesRowset classes.

Sample code: OpenTable4.zip

Moving around

A little change of plans is in order as instead of discussing bookmarks, I will take a deeper look into the IRowset::GetNextRows method. The reason behind this little change of plans is that I got fed up waiting for the sample's list view to populate. I told you before that this was the least-effective way of displaying large amounts of data in a list view: by filling it all up we have to wait for all the internal memory allocations to run and worse, we essentially duplicate the rowset data (much like reading a table into a .NET's DataSet). In fact, list views work much faster when they don't own the data but rely on the parent window to provide it for them. To do this, you must specify the LVS_OWNERDATA style bit when creating the list view control, thereby making it a virtual list view. Virtual list views query each item and sub-item data through the LVN_GETDISPINFO notification message (more info on this here). If you want to cache the data in advance, the virtual list view notifies the parent through an LVN_ODCACHEHINT notification, where the indexes of the rows about to be displayed are reported. This gives the parent a chance to cache the rows ahead. So how does this affect our sample code?

The first major difference os that we have to keep the CRowset object open thrughout the list view's lifetime so that ic can move back and forth to get the requested row. This means that we must find a way to map the zero-based list view item indexes and the table rows, and the most obvious one is to use the number of skipped rows from the rowset start. When the rowset is open, we must make sure that the internal row handle is initialized by calling MoveFirst. This will read the first row into memory and set the next reading position to "before the second row". What does this mean?

If you look at the IRowset::GetNextRows documentation, you see that the second (lRowsOffset) and third (cRows) parameters are signed numbers. We already know that the cRows parameter can only have the value of 1 in SQL Compact, because this engine only manages one row at a time. Interestingly it can also be -1, meaning that the cursor is reading backwards (the DBPROP_CANFETCHBACKWARDS is VARIANT_TRUE). This value is also by how much the "next read position" is incremented for next read. So if you open a brand-new rowset, call RestartPosition to position it "before the first row" and then consecutively call GetNextRows(NULL, 0, 1, ...) until the return value is different from S_OK, you will have visited all the rows from the rowset.

Moving back on the rowset would entail feeding a negative lRowsOffset value to the call but there are two gotchas:

1. What value would you use to read the previous row? If you think -1 you are wrong (if you keep cRows = 1) and the reason is simple: you would go back one row from the current reading position (after the current row) and put it before the current row, so you would effectively re-read the current row! To move back one row (with cRows = 1) you must set lRowsOffset to -2.
2. Can you really scroll back? If you look at all the available SQL Compact BOLs you will see that DBPROP_CANSCROLLBACKWARDS is read-only and VARIANT_FALSE, meaning that lRowsOffset cannot be negative. (I'm still trying to figure out if this is true for 3.0 and 3.5 becase elsewhere these BOLs say that you can have a negative offset.)

Just to be on the safe side, if you need to move back just restart the rowset position and move forwards from there (sigh).

Now go and have a look at the updated sample code. You will see some differences in the code, especially how fast the list view is displayed.

Download sample code: OpenTable3.zip