horse-performance-tuning

Horse Performance Tuning

To exploit the raw speed and low latency of the Horse framework, write handlers that avoid CPU bottlenecks and memory allocation overhead.

---

1. Minimizing Heap Allocations

Memory allocations (creating objects, large arrays, or concatenating strings) require thread synchronization locks in the memory manager, which slows down concurrent execution under heavy loads.

• Avoid Repeated JSON Parsing: If you just need to proxy or return static JSON payloads, send them as raw string strings or static stream resources rather than creating and destroying TJSONObject instances.
• Avoid String Concatenation: In loops, never concatenate strings using the + operator. Use TStringBuilder instead to avoid repeatedly reallocating memory on the heap.

// Inefficient (creates hundreds of temporary heap strings)
for I := 1 to 1000 do
  LResponseText := LResponseText + LData[I];

// Efficient
LBuilder := TStringBuilder.Create;
try
  for I := 1 to 1000 do
    LBuilder.Append(LData[I]);
  Res.Send(LBuilder.ToString);
finally
  LBuilder.Free;
end;

---

2. Fast Streaming for Large Payloads

When transferring large JSON strings, files, or reports, do not load the entire file contents into a string variable. Stream it directly to the socket chunk-by-chunk using Res.SendFile or Res.Download to maintain a low RAM profile.

• Bad: Loading a 100MB PDF into a TStringList or byte array.
• Good: Passing a TFileStream directly to the response (letting Horse stream it efficiently).

procedure ServeFileHandler(Req: THorseRequest; Res: THorseResponse);
var
  LStream: TFileStream;
begin
  LStream := TFileStream.Create('C:\data\largefile.zip', fmOpenRead or fmShareDenyWrite);
  Res.Status(THTTPStatus.OK).SendFile(LStream, 'largefile.zip');
  // Do NOT free LStream. Horse takes ownership of the stream.
end;

---

3. Selecting and Tuning the Transport Provider

The default Indy provider (Horse.Provider.Console) uses a thread-per-connection model. Under massive concurrency (thousands of connections), this model incurs thread-switching overhead.