Real-Time Processing

Streaming and online LOWESS for live data.

Overview

When data arrives continuously—from sensors, logs, or streaming pipelines—you need incremental smoothing that doesn't require reprocessing the entire dataset.


Online Mode: Point-by-Point

For true real-time applications where each point must be processed immediately.

window_capacity = 25 limits the internal buffer to the 25 most recent observations; each add_point call costs O(window) rather than growing with total history. min_points = 5 suppresses output until the window holds enough points for a stable fit — calls made before that threshold return null/None/nothing. update_mode = "incremental" re-fits only the most recent point rather than the full window, halving typical latency at a modest accuracy cost.

Sensor Data Example

using FastLOWESS

# Simulate sensor readings
times = collect(Float64, 0:99)
temperatures = 20.0 .+ 5.0 .* sin.(times ./ 10.0) .+ sin.(times .* 1.7) .* 0.5

# Process with online mode
model = OnlineLowess(;
    fraction=0.3,
    window_capacity=25,
    min_points=5,
    update_mode="incremental"
)
count = 0
for i in eachindex(times)
    result = add_point(model, times[i], temperatures[i])
    if result !== nothing
        if count < 5
            println("Time $(times[i]): smoothed = $(round(result.y; digits=4))")
        end
        global count += 1
    end
end
println("... ($(count - 5) more)")
Time 4.0: smoothed = 22.1941
Time 5.0: smoothed = 22.7964
Time 6.0: smoothed = 22.4733
Time 7.0: smoothed = 22.912
Time 8.0: smoothed = 24.0164
... (91 more)

Streaming Mode: Chunk Processing

For large datasets that arrive in batches or files.

chunk_size controls how many data points are processed in one pass; matching it to your file-read buffer or message-batch size avoids unnecessary copying. overlap retains that many points from the previous chunk as context so the neighbourhood at chunk boundaries is not artificially truncated. merge_strategy = "weighted_average" blends the overlapping region smoothly; use "last" if chunk boundaries are guaranteed to be well separated and no blending is needed.

Always call finalize()

The streaming adapter buffers overlap data. Call finalize() after the last chunk to retrieve the buffered tail.

Log File Processing

using FastLOWESS

chunk1_x = collect(Float64, 0:49)
chunk1_y = sin.(chunk1_x) .+ 0.1
chunk2_x = collect(Float64, 50:99)
chunk2_y = sin.(chunk2_x) .+ 0.1

# Streaming mode handles everything internally
model = StreamingLowess(;
    fraction=0.1,
    iterations=2,
    chunk_size=50,
    overlap=10,
    merge_strategy="weighted_average"
)
process_chunk(model, chunk1_x, chunk1_y)
process_chunk(model, chunk2_x, chunk2_y)
result = finalize(model)
println("y[0]: ", result.y[1])
y[0]: 0.518852654979974

Real-Time Dashboard Example

The dashboard pattern uses a plain LOWESS fit on a manually managed sliding window rather than OnlineLowess. This is the simplest approach when your UI framework already owns the data buffer and you only need the most recent smoothed value per frame. The trade-off is a full O(window^2) refit on every tick; for high-frequency streams prefer OnlineLowess with update_mode = "incremental" to bound per-frame cost.

using FastLOWESS

n = 100
x = collect(range(0, 2π, length=n))
y = sin.(x) .+ 0.1

window_x = Float64[]
window_y = Float64[]
latest = 0.0
for i in eachindex(x)
    push!(window_x, x[i])
    push!(window_y, y[i])
    if length(window_x) > 50
        popfirst!(window_x)
        popfirst!(window_y)
    end
    length(window_x) < 2 && continue

    model = Lowess(; fraction=0.4)
    result = fit(model, window_x, window_y)
    global latest = result.y[end]
end
println("Smoothed (dashboard, latest tick): ", latest)
Smoothed (dashboard, latest tick): -0.06780546949749158

Choosing Parameters

Online Mode

ParameterGuidance
window_capacityEnough history for fraction to work
min_points2–5 typically; higher for stability
update_mode"incremental" for speed, "full" for accuracy

Streaming Mode

ParameterGuidance
chunk_sizeBalance memory vs. processing overhead
overlap10–20% of chunk_size for smooth transitions
merge_strategy"weighted_average" for best quality, "average" for simplicity

Performance Considerations

ModeMemoryLatencyUse Case
OnlineFixed (window)~1ms/pointSensors, dashboards
Streaming~chunk_size~100ms/chunkLarge files, ETL
BatchFull datasetN/AAnalysis, reports

See Also