Load, Replicas & Resilience · Lesson 2

Latency distributions and Little’s Law

Keep percentile definitions and steady-state units explicit.

External .NET10 SDK10.0.403, real Kestrel and HTTP driver. Experiments are limited to owned loopback processes. SQLite10.0.14 is a local-file classroom ledger, not a network production database.

Read or download the complete SDK project, then follow its setup on your computer. The separate C# companion practice below runs in SharpRune’s .NET 9 console environment.

A percentile describes a selected collection of observations

Percentiles describe a distribution only after the observation population and estimator are defined. The teaching code uses nearest rank over five sorted latency values. The real driver applies that estimator to recorded HTTP attempt durations, including attempts that receive rejection responses. Those are not automatically the same population as successful logical operations or an end-to-end operation containing retries and waits. Before comparing p95 values, identify which events entered each collection. The portable result is deterministic arithmetic; the external values come from Stopwatch measurements and vary with the actual process, workload and environment.

Calculate the ranks rather than guessing an average

The fixed observations are ten, twenty, thirty, forty and one hundred. Nearest rank selects the ceiling of probability times sample count, then uses that one-based rank in the sorted array. For p50 the rank is three, yielding thirty. For p95 the rank is five, yielding one hundred. The final console line divides average in-system work four by average time zero point two seconds and prints rate:20.00. This is a units-explicit Little’s Law illustration under a stable boundary, not measured capacity or permission to offer twenty requests each second to an untested service.

Use consistent boundaries when applying Little’s Law

The relationship between average in-system work, average rate and average time requires those averages to refer to the same system and a suitable stable observation interval. Including queue time in one measurement but counting only workers in another breaks that interpretation. A growing queue is evidence that a simple steady-state capacity claim needs more care. Likewise, substituting p95 latency for average time changes the quantity in the calculation; it is not the same law. Use the formula to check units and reason about observed occupancy, not to turn an intended client arrival rate into a guarantee of sustainable throughput.

Tail estimates are coarse in a bounded classroom run

With five samples, both p95 and p99 select the maximum under nearest rank. A sixteen-request driver run remains a small sample, so a single slow startup or scheduling event can strongly affect its tail. Percentile ordering is a useful sanity check, but precise tail guarantees need a deliberate measurement design and more evidence than this bounded exercise. The driver reports counts alongside its p50, p95 and p99 so that limitation stays visible. Avoid averaging replica percentiles into a supposed global percentile; the underlying distributions and request weights determine the combined population.

Practice: change one tail observation

Replace only the fixed value one hundred with one thousand in the portable companion. Predict p50 and p95 before running, and explain why the median can remain unchanged while the tail changes dramatically. Then replace the fixed sample with only two observations and compute nearest-rank p95 manually. This exercise checks the estimator rather than a preferred intuition about smooth percentiles. Keep the Little’s Law example unchanged so a tail-data edit cannot silently change an independently supplied average-time assumption. The hint is to calculate the rank first and only then select the value.

Practice: name the units and include queue time consistently

Write the rate example once with time in seconds and once with the same time expressed as two hundred milliseconds. Convert units explicitly and confirm the same final requests-per-second value. Then describe a queueing observation boundary in which average work includes both active and waiting operations. State which elapsed interval a corresponding average time must include. Do not assert a capacity result from those descriptions alone. The independent objective is to make every unit and boundary visible enough that another reviewer can identify a mismatched numerator or an omitted waiting phase.

Preserve distributions and counts when comparing runs

The external driver can provide measured attempts, outcomes and duration percentiles for a bounded local experiment. Keep those values together and record whether retries or rejections entered the latency population. The console establishes exact nearest-rank examples and one arithmetic average relationship, while real process behavior supplies different evidence. A useful review asks whether the sample is representative, whether the generator skipped demand, whether the queue grew and whether the compared populations match. That review is more informative than reporting one impressive p95 without the settings and counts needed to understand how it was obtained.

  • Define percentile estimator and observation population before interpreting tail latency.
  • Use Little’s Law with averages, consistent units and a stated stable system boundary.
External .NET10 SDK10.0.403, real Kestrel and HTTP driver. Experiments are limited to owned loopback processes. SQLite10.0.14 is a local-file classroom ledger, not a network production database.

Latency distributions and Little’s Law — real bounded loopback testbed

Set up the SDK project

  1. Save the complete Server and Driver directories plus global.json into one new owned folder. Install official .NET10 SDK10.0.403.
  2. Restore both projects explicitly: dotnet restore Server/Lab.Server.csproj --source https://api.nuget.org/v3/index.json ; then dotnet restore Driver/Lab.Driver.csproj --source https://api.nuget.org/v3/index.json
  3. In terminal1 start the first owned server: ASPNETCORE_ENVIRONMENT=Development dotnet run --project Server/Lab.Server.csproj --no-launch-profile --urls http://127.0.0.1:5081 --Lab:Replica=one --Lab:LedgerPath=./owned-ledger.db --Lab:Workers=2 --Lab:QueueCapacity=4 --Lab:PoolSize=2 --Lab:WorkMilliseconds=20
  4. For two-replica experiments use terminal2 with port5082, Lab:Replica=two and Lab:Partition=1. Resolve both ledger paths to the same owned absolute file only when the lesson explicitly tests shared local coordination.
  5. Run drivers only against your owned loopback ports. Maximum32 logical operations, maximum8 in flight and at most2 retry attempts keep each run below100 requests. Outputs include variable measured timings.
  6. Stop only the processes you started using Ctrl+C in their own terminals; clean up only your owned fixture directory.
  7. Run the closed driver with16 operations and inspect its nearest-rank p50/p95/p99 plus attempted/success counts. Repeat once without asserting the same timing values.

Download complete SDK project ↗

Project files

global.jsonDownload file ↗
{
  "sdk": {
    "version": "10.0.403",
    "rollForward": "latestPatch"
  }
}
Server/Lab.Server.csprojDownload file ↗
<Project Sdk="Microsoft.NET.Sdk.Web"><PropertyGroup><TargetFramework>net10.0</TargetFramework><Nullable>enable</Nullable><ImplicitUsings>enable</ImplicitUsings></PropertyGroup><ItemGroup><PackageReference Include="Microsoft.Data.Sqlite" Version="10.0.14" /></ItemGroup></Project>
Server/Program.csDownload file ↗
using Microsoft.AspNetCore.Hosting.Server;
using Microsoft.AspNetCore.Hosting.Server.Features;
using Microsoft.Data.Sqlite;
var builder = WebApplication.CreateBuilder(args);
var settings = new LabSettings {
    Replica = builder.Configuration["Lab:Replica"] ?? "one",
    Workers = builder.Configuration.GetValue("Lab:Workers", 2),
    Partition = builder.Configuration.GetValue("Lab:Partition", 0),
    QueueCapacity = builder.Configuration.GetValue("Lab:QueueCapacity", 4),
    PoolSize = builder.Configuration.GetValue("Lab:PoolSize", 2),
    WorkMilliseconds = builder.Configuration.GetValue("Lab:WorkMilliseconds", 20),
    LedgerPath = builder.Configuration["Lab:LedgerPath"] ?? Path.Combine(Path.GetTempPath(), "sharprune-local-ledger.db")
};
settings.Validate(); builder.Services.AddSingleton(settings); builder.Services.AddSingleton<LabRuntime>(); builder.Services.AddSingleton<LocalIo>();
builder.Services.AddSingleton<IHostedService>(services => services.GetRequiredService<LabRuntime>());
builder.Services.AddSingleton<LocalLedger>(); var app = builder.Build();
app.Services.GetRequiredService<LocalLedger>().Initialize();
app.MapGet("/instance", (LabRuntime lab) => new { replica = settings.Replica, visits = lab.Visit() });
app.MapGet("/live", () => TypedResults.Text("live"));
app.MapGet("/io-delay", async (CancellationToken token) => { await Task.Delay(20, token); return TypedResults.Text("controlled loopback reply"); });
app.MapGet("/io", async (CancellationToken token, IServer server, LocalIo io) => {
    using var response = await io.Client.GetAsync(LocalAddress(server), token); response.EnsureSuccessStatusCode(); _ = await response.Content.ReadAsByteArrayAsync(token);
    return TypedResults.Ok(new { replica = settings.Replica, kind = "async-loopback-http" });
});
app.MapGet("/sync", (CancellationToken token, IServer server, LocalIo io) => {
    using var response = io.Client.GetAsync(LocalAddress(server), token).GetAwaiter().GetResult(); response.EnsureSuccessStatusCode(); _ = response.Content.ReadAsByteArrayAsync(token).GetAwaiter().GetResult();
    return TypedResults.Ok(new { replica = settings.Replica, kind = "deliberate-blocking-demo" });
});
app.MapGet("/cpu", () => { uint value = 1; for (int i = 0; i < 200000; i++) value = unchecked(value * 1664525 + 1013904223); return TypedResults.Ok(new { replica = settings.Replica, kind = "bounded-cpu", value }); });
app.MapGet("/threadpool", () => new { threads = ThreadPool.ThreadCount, pending = ThreadPool.PendingWorkItemCount, completed = ThreadPool.CompletedWorkItemCount });
app.MapGet("/ready", (LabRuntime lab) => lab.Accepting ? Results.Ok(new { ready = true }) : Results.StatusCode(503));
app.MapGet("/stats", (LabRuntime lab) => new { replica = settings.Replica, accepting = lab.Accepting, main = lab.Main.Snapshot(), batch = lab.Batch.Snapshot(), dependencyCalls = lab.Circuit.DependencyCalls });
app.MapGet("/work", (HttpContext context, LabRuntime lab, int? budgetMs) => Handle(context, lab.Main, lab, budgetMs));
app.MapGet("/batch", (HttpContext context, LabRuntime lab, int? budgetMs) => Handle(context, lab.Batch, lab, budgetMs));
app.MapGet("/unstable", (LabRuntime lab, string? key) => {
    if (key != "demo" && key != "always") return Results.BadRequest();
    int attempt = lab.Attempt(key); return key == "always" || attempt == 1 ? Results.StatusCode(503) : Results.Ok(new { replica = settings.Replica, attempt });
});
app.MapGet("/session/{name}", (LabRuntime lab, string name) => name == "demo" ? Results.Ok(new { replica = settings.Replica, value = lab.Session() }) : Results.BadRequest());
app.MapPost("/once/{key}", (string key, LedgerInput input, LocalLedger ledger) => {
    if (key != "demo" || input.Amount is < 1 or > 100) return Results.BadRequest();
    LedgerDecision result = ledger.Record(key, input.Amount); return result.Conflict ? Results.Conflict(new { error = "key amount conflict" }) : Results.Ok(result);
});
app.MapGet("/ledger", (LocalLedger ledger) => new { receipts = ledger.Count() });
app.MapGet("/guarded", (LabRuntime lab) => lab.Circuit.Invoke(lab.DependencyFailing));
app.MapGet("/partition/{key}", (string key, int? owner) => {
    if (key != "A" && key != "B") return Results.BadRequest();
    int partition = StablePartition.For(key, 2);
    return owner == partition && settings.Partition == partition ? Results.Ok(new { replica = settings.Replica, partition, key }) : Results.Conflict(new { expectedPartition = partition });
});
if (app.Environment.IsDevelopment()) {
    app.MapPost("/lab/gate/{lane}/{action}", (string lane, string action, LabRuntime lab) => {
        WorkLane? target = lane == "main" ? lab.Main : lane == "batch" ? lab.Batch : null;
        if (target is null || action != "open" && action != "close") return Results.BadRequest();
        if (action == "open") target.Gate.Open(); else target.Gate.Close(); return Results.NoContent();
    });
    app.MapPost("/lab/drain", (LabRuntime lab) => { lab.Drain(); return Results.NoContent(); });
    app.MapPost("/lab/dependency/{failing:bool}", (bool failing, LabRuntime lab) => { lab.DependencyFailing = failing; return Results.NoContent(); });
}
app.Run();
static async Task<IResult> Handle(HttpContext context, WorkLane lane, LabRuntime lab, int? budgetMs) {
    if (budgetMs is < 1 or > 1000) return Results.BadRequest();
    using var budget = CancellationTokenSource.CreateLinkedTokenSource(context.RequestAborted);
    if (budgetMs is int milliseconds) budget.CancelAfter(milliseconds);
    var item = new WorkItem(budget.Token);
    AdmissionDecision decision = lab.Admit(lane, item);
    if (decision == AdmissionDecision.Draining) return Results.StatusCode(503);
    if (decision == AdmissionDecision.Full) return Results.StatusCode(429);
    try { await item.Completion.Task.WaitAsync(budget.Token); return Results.Ok(new { replica = lab.Settings.Replica, completed = true }); }
    catch (OperationCanceledException) when (!context.RequestAborted.IsCancellationRequested) { return Results.Problem(statusCode: 504, title: "Classroom work budget expired"); }
}
static Uri LocalAddress(IServer server) {
    string address = server.Features.Get<IServerAddressesFeature>()!.Addresses.Single();
    return new UriBuilder("http", "127.0.0.1", new Uri(address).Port) { Path = "/io-delay" }.Uri;
}
public sealed class LocalIo : IDisposable { public HttpClient Client { get; } = new() { Timeout = TimeSpan.FromSeconds(5) }; public void Dispose() => Client.Dispose(); }
public sealed record LedgerInput(int Amount);
public sealed class LabSettings {
    public string Replica { get; init; } = "one"; public int Partition { get; init; } public int Workers { get; init; } public int QueueCapacity { get; init; }
    public int PoolSize { get; init; } public int WorkMilliseconds { get; init; } public string LedgerPath { get; init; } = "";
    public void Validate() {
        if (Partition is < 0 or > 1 || Replica.Length is < 1 or > 32 || Workers is < 1 or > 8 || QueueCapacity is < 1 or > 16 || PoolSize is < 1 or > 8 || WorkMilliseconds is < 1 or > 250 || string.IsNullOrWhiteSpace(LedgerPath)) throw new InvalidOperationException("Invalid bounded lab settings");
    }
}
public enum AdmissionDecision { Accepted, Full, Draining }
public sealed class LabRuntime : IHostedService, IDisposable {
    private readonly object admissionSync = new();
    private int accepting = 1, visits, sessions, demoAttempts, alwaysAttempts, failing = 1, disposed;
    public LabSettings Settings { get; } public WorkLane Main { get; } public WorkLane Batch { get; }
    public ClassroomCircuit Circuit { get; } = new();
    public LabRuntime(LabSettings settings) { Settings = settings; Main = new(settings.Workers, settings.QueueCapacity, settings.PoolSize, settings.WorkMilliseconds); Batch = new(1, 1, 1, settings.WorkMilliseconds); }
    public bool Accepting => Volatile.Read(ref accepting) == 1;
    public bool DependencyFailing { get => Volatile.Read(ref failing) == 1; set => Volatile.Write(ref failing, value ? 1 : 0); }
    public void Drain() { lock (admissionSync) Volatile.Write(ref accepting, 0); }
    public AdmissionDecision Admit(WorkLane lane, WorkItem item) {
        lock (admissionSync) {
            if (!Accepting) return AdmissionDecision.Draining;
            return lane.TrySubmit(item) ? AdmissionDecision.Accepted : AdmissionDecision.Full;
        }
    }
    public int Visit() => Interlocked.Increment(ref visits); public int Session() => Interlocked.Increment(ref sessions);
    public int Attempt(string key) => key == "demo" ? Interlocked.Increment(ref demoAttempts) : Interlocked.Increment(ref alwaysAttempts);
    public Task StartAsync(CancellationToken token) { Main.Start(); Batch.Start(); return Task.CompletedTask; }
    public Task StopAsync(CancellationToken token) => Task.WhenAll(Main.StopAsync(token), Batch.StopAsync(token));
    public void Dispose() { if (Interlocked.Exchange(ref disposed, 1) != 0) return; Main.Dispose(); Batch.Dispose(); }
}
public static class StablePartition {
    public static int For(string key, int count) { uint hash = 2166136261; foreach (byte b in System.Text.Encoding.UTF8.GetBytes(key)) hash = unchecked((hash ^ b) * 16777619); return (int)(hash % (uint)count); }
}
Server/WorkLane.csDownload file ↗
using System.Threading.Channels;
public sealed record WorkItem(CancellationToken Token) { public TaskCompletionSource<bool> Completion { get; } = new(TaskCreationOptions.RunContinuationsAsynchronously); }
public sealed class AsyncGate {
    private readonly object sync = new(); private TaskCompletionSource<bool> source = Completed();
    private static TaskCompletionSource<bool> Completed() { var value = new TaskCompletionSource<bool>(TaskCreationOptions.RunContinuationsAsynchronously); value.SetResult(true); return value; }
    public void Close() { lock (sync) { if (source.Task.IsCompleted) source = new(TaskCreationOptions.RunContinuationsAsynchronously); } }
    public void Open() { lock (sync) source.TrySetResult(true); }
    public Task WaitAsync(CancellationToken token) { lock (sync) return source.Task.WaitAsync(token); }
}
public sealed class WorkLane : IDisposable {
    private readonly Channel<WorkItem> queue; private readonly SemaphoreSlim pool; private readonly int workers, milliseconds;
    private readonly CancellationTokenSource stop = new(); private Task[] running = Array.Empty<Task>();
    private int active, inPool, maxActive, maxPool, accepted, rejected, completed, canceled, disposed;
    public AsyncGate Gate { get; } = new();
    public WorkLane(int workers, int capacity, int poolSize, int milliseconds) {
        this.workers = workers; this.milliseconds = milliseconds; pool = new(poolSize, poolSize);
        queue = Channel.CreateBounded<WorkItem>(new BoundedChannelOptions(capacity) { FullMode = BoundedChannelFullMode.Wait, SingleWriter = false, SingleReader = workers == 1 });
    }
    public void Start() { running = Enumerable.Range(0, workers).Select(_ => RunAsync()).ToArray(); }
    public bool TrySubmit(WorkItem item) { if (queue.Writer.TryWrite(item)) { Interlocked.Increment(ref accepted); return true; } Interlocked.Increment(ref rejected); return false; }
    private async Task RunAsync() {
        await foreach (WorkItem item in queue.Reader.ReadAllAsync(stop.Token)) {
            int now = Interlocked.Increment(ref active); Maximum(ref maxActive, now); bool entered = false;
            using var lifetime = CancellationTokenSource.CreateLinkedTokenSource(item.Token, stop.Token);
            try {
                lifetime.Token.ThrowIfCancellationRequested(); await Gate.WaitAsync(lifetime.Token);
                await pool.WaitAsync(lifetime.Token); entered = true; Maximum(ref maxPool, Interlocked.Increment(ref inPool));
                await Task.Delay(milliseconds, lifetime.Token); Interlocked.Increment(ref completed); item.Completion.TrySetResult(true);
            } catch (OperationCanceledException) { Interlocked.Increment(ref canceled); item.Completion.TrySetCanceled(lifetime.Token); }
            finally { if (entered) { Interlocked.Decrement(ref inPool); pool.Release(); } Interlocked.Decrement(ref active); }
        }
    }
    private static void Maximum(ref int field, int value) { int old; do { old = Volatile.Read(ref field); if (old >= value) return; } while (Interlocked.CompareExchange(ref field, value, old) != old); }
    public object Snapshot() => new { active = Volatile.Read(ref active), queued = queue.Reader.Count, inPool = Volatile.Read(ref inPool), maxActive = Volatile.Read(ref maxActive), maxPool = Volatile.Read(ref maxPool), accepted = Volatile.Read(ref accepted), rejected = Volatile.Read(ref rejected), completed = Volatile.Read(ref completed), canceled = Volatile.Read(ref canceled) };
    public async Task StopAsync(CancellationToken token) {
        queue.Writer.TryComplete();
        try { await Task.WhenAll(running).WaitAsync(token); }
        catch (OperationCanceledException) when (token.IsCancellationRequested) { stop.Cancel(); Gate.Open(); try { await Task.WhenAll(running); } catch (OperationCanceledException) { } }
    }
    public void Dispose() { if (Interlocked.Exchange(ref disposed, 1) != 0) return; stop.Cancel(); pool.Dispose(); stop.Dispose(); }
}
Server/LocalLedger.csDownload file ↗
using Microsoft.Data.Sqlite;
public sealed record LedgerDecision(bool Replay, bool Conflict, int Amount);
public sealed class LocalLedger {
    private readonly string connectionString;
    public LocalLedger(LabSettings settings) { connectionString = new SqliteConnectionStringBuilder { DataSource = settings.LedgerPath, DefaultTimeout = 2 }.ToString(); }
    private SqliteConnection Open() { var connection = new SqliteConnection(connectionString); connection.Open(); return connection; }
    public void Initialize() { using var connection = Open(); using var command = connection.CreateCommand(); command.CommandText = "CREATE TABLE IF NOT EXISTS receipts (request_key TEXT PRIMARY KEY, amount INTEGER NOT NULL)"; command.ExecuteNonQuery(); }
    public LedgerDecision Record(string key, int amount) {
        using var connection = Open(); using var transaction = connection.BeginTransaction();
        using var insert = connection.CreateCommand(); insert.Transaction = transaction;
        insert.CommandText = "INSERT INTO receipts(request_key, amount) VALUES ($key, $amount) ON CONFLICT(request_key) DO NOTHING";
        insert.Parameters.AddWithValue("$key", key); insert.Parameters.AddWithValue("$amount", amount); int changed = insert.ExecuteNonQuery();
        using var read = connection.CreateCommand(); read.Transaction = transaction; read.CommandText = "SELECT amount FROM receipts WHERE request_key = $key"; read.Parameters.AddWithValue("$key", key);
        int stored = Convert.ToInt32(read.ExecuteScalar(), System.Globalization.CultureInfo.InvariantCulture); transaction.Commit();
        return new LedgerDecision(changed == 0, stored != amount, stored);
    }
    public int Count() { using var connection = Open(); using var read = connection.CreateCommand(); read.CommandText = "SELECT COUNT(*) FROM receipts"; return Convert.ToInt32(read.ExecuteScalar(), System.Globalization.CultureInfo.InvariantCulture); }
}
Server/ClassroomCircuit.csDownload file ↗
public sealed class ClassroomCircuit {
    private readonly object sync = new(); private int failures, calls; private bool probe; private DateTimeOffset openUntil = DateTimeOffset.MinValue;
    public int DependencyCalls { get { lock (sync) return calls; } }
    public IResult Invoke(bool failing) {
        lock (sync) {
            DateTimeOffset now = DateTimeOffset.UtcNow;
            if (openUntil != DateTimeOffset.MinValue && now < openUntil) return Results.Json(new { state = "open", called = false }, statusCode: 503);
            if (probe) return Results.Json(new { state = "probe-busy", called = false }, statusCode: 503);
            bool halfOpen = openUntil != DateTimeOffset.MinValue; if (halfOpen) probe = true;
            calls++;
            // The dependency is an explicitly controlled in-process classroom outcome, not a real remote service.
            if (failing) {
                failures++; if (halfOpen || failures >= 2) openUntil = now.AddMilliseconds(250); probe = false;
                return Results.Json(new { state = openUntil == DateTimeOffset.MinValue ? "closed" : "open", called = true }, statusCode: 503);
            }
            failures = 0; openUntil = DateTimeOffset.MinValue; probe = false; return Results.Ok(new { state = "closed", called = true });
        }
    }
}
Driver/Lab.Driver.csprojDownload file ↗
<Project Sdk="Microsoft.NET.Sdk"><PropertyGroup><OutputType>Exe</OutputType><TargetFramework>net10.0</TargetFramework><Nullable>enable</Nullable><ImplicitUsings>enable</ImplicitUsings></PropertyGroup></Project>
Driver/Program.csDownload file ↗
using System.Collections.Concurrent;
using System.Diagnostics;
using System.Globalization;
using System.Text.Json;
if (args.Length < 2) throw new ArgumentException("Usage: closed|open|retry|shard|hot|io|cpu|sync baseUrl[,baseUrl] [count<=32] [concurrency<=8] [intervalMs]");
string mode = args[0];
if (!new[] { "closed", "open", "retry", "shard", "hot", "io", "cpu", "sync" }.Contains(mode)) throw new ArgumentException("Unknown mode");
Uri[] targets = args[1].Split(',').Select(s => new Uri(s.TrimEnd('/') + "/")).ToArray();
if (targets.Length is < 1 or > 2 || targets.Any(u => !u.IsLoopback || u.Scheme != "http" || u.AbsolutePath != "/" || !string.IsNullOrEmpty(u.Query) || !string.IsNullOrEmpty(u.UserInfo))) throw new ArgumentException("Only one or two owned HTTP loopback roots are allowed");
int count = args.Length > 2 ? int.Parse(args[2], CultureInfo.InvariantCulture) : 16;
int concurrency = args.Length > 3 ? int.Parse(args[3], CultureInfo.InvariantCulture) : 4;
int interval = args.Length > 4 ? int.Parse(args[4], CultureInfo.InvariantCulture) : 2;
if (count is < 1 or > 32 || concurrency is < 1 or > 8 || interval is < 1 or > 100 || (mode == "shard" || mode == "hot") && targets.Length != 2) throw new ArgumentException("Bounded local settings required");
using var client = new HttpClient { Timeout = TimeSpan.FromSeconds(5) };
var latencies = new ConcurrentBag<double>(); var replicas = new ConcurrentDictionary<string, byte>(StringComparer.Ordinal);
int attempted = 0, successful = 0, rejected = 0, skipped = 0, logical = 0;
async Task Invoke(int index) {
    string key = mode == "hot" || index % 2 == 0 ? "A" : "B"; int partition = Partition(key);
    Uri target = targets[(mode == "shard" || mode == "hot") ? partition : index % targets.Length];
    string route = mode switch { "retry" => "unstable?key=demo", "shard" or "hot" => "partition/" + key + "?owner=" + partition, "io" => "io", "cpu" => "cpu", "sync" => "sync", _ => "work" };
    int maximumAttempts = mode == "retry" ? 2 : 1;
    for (int attempt = 1; attempt <= maximumAttempts; attempt++) {
        var begin = Stopwatch.GetTimestamp(); Interlocked.Increment(ref attempted);
        using var request = new HttpRequestMessage(HttpMethod.Get, new Uri(target, route));
        using var response = await client.SendAsync(request); string body = await response.Content.ReadAsStringAsync();
        latencies.Add(Stopwatch.GetElapsedTime(begin).TotalMilliseconds);
        if ((int)response.StatusCode == 503 && attempt < maximumAttempts) { await Task.Delay(10 + index % 7); continue; }
        if (response.IsSuccessStatusCode) {
            Interlocked.Increment(ref successful); using var doc = JsonDocument.Parse(body);
            if (doc.RootElement.TryGetProperty("replica", out var name)) replicas.TryAdd(name.GetString()!, 0);
        } else Interlocked.Increment(ref rejected);
        break;
    }
}
var elapsed = Stopwatch.StartNew();
if (mode == "open") {
    using var slots = new SemaphoreSlim(concurrency, concurrency); var running = new List<Task>();
    for (int i = 0; i < count; i++) {
        double wait = i * interval - elapsed.Elapsed.TotalMilliseconds; if (wait > 0) await Task.Delay(TimeSpan.FromMilliseconds(wait));
        Interlocked.Increment(ref logical);
        if (!slots.Wait(0)) { Interlocked.Increment(ref skipped); continue; }
        int index = i; running.Add(Release(index));
    }
    await Task.WhenAll(running);
    async Task Release(int index) { try { await Invoke(index); } finally { slots.Release(); } }
} else {
    int next = -1;
    await Task.WhenAll(Enumerable.Range(0, concurrency).Select(async _ => {
        while (true) { int index = Interlocked.Increment(ref next); if (index >= count) return; Interlocked.Increment(ref logical); await Invoke(index); }
    }));
}
elapsed.Stop(); double[] sorted = latencies.OrderBy(x => x).ToArray();
double Percentile(double probability) => sorted.Length == 0 ? 0 : sorted[(int)Math.Ceiling(probability * sorted.Length) - 1];
Console.WriteLine(JsonSerializer.Serialize(new { mode, scheduled = logical, attempted, successful, rejected, skipped, concurrency, elapsedMs = elapsed.Elapsed.TotalMilliseconds, p50Ms = Percentile(.5), p95Ms = Percentile(.95), p99Ms = Percentile(.99), replicas = replicas.Keys.OrderBy(x => x).ToArray() }));
static int Partition(string key) { uint hash = 2166136261; foreach (byte b in System.Text.Encoding.UTF8.GetBytes(key)) hash = unchecked((hash ^ b) * 16777619); return (int)(hash % 2); }
Server/appsettings.jsonDownload file ↗
{
  "Lab": {
    "Replica": "one",
    "Workers": 2,
    "QueueCapacity": 4,
    "PoolSize": 2,
    "WorkMilliseconds": 20
  }
}

Expected result

Actual durations are nonnegative measured values with p50<=p95<=p99. Counts and settings are reproducible; exact latency and throughput are not.

Use the listed SDK and dependencies for this project. Open C# practice runs the separate portable exercise.

C# companion practice

using System;
using System.Linq;
using System.Collections.Generic;
using System.Globalization;
using System.Threading;
using System.Threading.Tasks;

int[] sorted = new[] { 10, 20, 30, 40, 100 }; int P(double p) => sorted[(int)Math.Ceiling(p * sorted.Length) - 1];
Console.WriteLine("p50:" + P(.5)); Console.WriteLine("p95:" + P(.95)); Console.WriteLine("rate:" + (4m / .2m).ToString("0.00", CultureInfo.InvariantCulture));
Open C# practice

C# practice output

p50:30
p95:100
rate:20.00
Try this
Define the population, percentile estimator, units and stable boundary before applying L = rate × time.

Further reading