Synopse Open Source

For real applications, retrieving objects per ID is not enough.
Your project may have the need to retrieve objects by a textual field, e.g. a name or identifier.
In this case, you can specify a published property of the TSQLRecord as stored false, and it will be defined as an unique column in the underlying database.

For instance, in the latest version of our performance benchmark sample code, you can define the UNIK conditional to define both LastName and FirstName properties as unique:

type
  TSQLRecordSample = class(TSQLRecord)
  private
    fFirstName: RawUTF8;
    fLastName: RawUTF8;
    fAmount: currency;
    fBirthDate: TDateTime;
    fLastChange: TModTime;
    fCreatedAt: TCreateTime;
  published
    property FirstName: RawUTF8 index 40 read fFirstName write fFirstName
      {$ifdef UNIK}stored false{$endif};
    property LastName: RawUTF8 index 40 read fLastName write fLastName
      {$ifdef UNIK}stored false{$endif};
    property Amount: currency read fAmount write fAmount;
    property BirthDate: TDateTime read fBirthDate write fBirthDate;
    property LastChange: TModTime read fLastChange;
    property CreatedAt: TCreateTime read fCreatedAt write fCreatedAt;
  end;

During insertion or update of records, the database will have to check for uniqueness of those column values. It will have an additional performance cost, since a search of the new value is to be performed among existing values.
In order to speed-up the process, a so-called index is created at the database level.
As a consequence, further lookup using this property will benefit for this index, and will be much faster than a classic loop throughout all data.

In the mORMot core, we just made some modifications related to this feature:

• External tables are now able to create properly UNIQUE fields at database level, as expected;
• A hashed-based index feature has been added to the fast TSQLRestServerStaticInMemory internal engine, and insertion will have no speed penalty any more.

After having tested and enhanced the external database speed (including BATCH mode), we are now able to benchmark all database engines available in mORMot.

In fact, the ORM part of our framework has several potential database backends, in addition to the default SQLite3 file-based engine.
Each engine may have its own purpose, according to the application expectation.

The following tables try to sum up all available possibilities, and give some benchmark (average rows/seconds for writing or read). 

In these tables:

• 'internal' means use of the internal SQLite3 engine;
• 'external' stands for an external access via SynDB;
• 'TObjectList' indicates a TSQLRestServerStaticInMemory instance either static (with no SQL support) or virtual (i.e. SQL featured via SQLite3 virtual table mechanism) which may persist the data on disk as JSON or compressed binary;
• 'trans' stands for Transaction, i.e. when the write process is nested within BeginTransaction / Commit calls;
• 'batch' mode will be described in this article;
• 'read one' states that one object is read per call (ORM generates a SELECT * FROM table WHERE ID=?);
• 'read all' is when all 5000 objects are read in a single call (i.e. running SELECT * FROM table);
• ACID is an acronym for "Atomicity Consistency Isolation Durability" properties, which guarantee that database transactions are processed reliably: for instance, in case of a power loss or hardware failure, the data will be saved on disk in a consistent way, with no potential loss of data.

A common issue with Client-Server databases is the latency introduced for each query.

For example, suppose you have a requirement to first collect some information from your application’s users and then insert that information into a table in Oracle Database.

The first obvious option is to insert these multiple rows into the table through a loop in your program. This loop iterates over the data to be inserted and does what is known as a single-row insert , because the application sends one single row of data to the database at a time. Due to the network latency (typically around 1 ms over a corporate network), it would achieve not more than 500-600 requests per second to let the work done, since for each INSERT, a so-called round-trip occurs: a message is sent to Oracle, then a response is sent back to the client.

You have another option for inserting multiple rows of data into the table— that reduces the number of round-trips and improves application performance, database performance, and network resource use. Rather than having the application send a single row of data to the database at a time, it can use array binding to send the data in batches of rows. Therefore, you reduce a lot the number of round-trips to be processed, and enhance performance by a factor of about 100.

Our SynDB unit has been enhanced to introduce new TSQLDBStatement.BindArray() methods, introducing array binding for faster database batch modifications (only implemented in SynDBOracle by now - but MS SQL has a similar feature called OleDB bulk insert).
It is available from the ORM side or mORMot, when working with external tables, in BATCH mode.

Thanks to this enhancement, inserting records within Oracle comes from 400-500 rows per second to more than 50000 rows per second!
It was also a good opportunity to speed up the BATCH process globally, and to benchmark our Oracle back-end against existing external databases, i.e. SQLite3 (as file or in-memory), and Jet / MS Access / .mdb engine.

Note that this article scope is only about virtual tables linked to external databases (i.e. TSQLRecordExternal). Plain TSQLRecord classes will access directly to the SQLite3 engine or in-memory TList, so speed will be even higher than the below values.

Featuring benchmark source code and nice performance charts.