tracing-console

An interactive TUI for inspecting tracing spans coming off a live server, plus the cache and RPC layer that gets them to you.

The tracing tie-in, tracing-cache, is made to have the lowest possible overhead while disabled. It's expected that you will have nothing connected to your tracing port almost all the time. In the rare case you need it, you can connect, enable briefly, (or for longer with a non-100% sampling rate) and then disable. In this way you can capture detailed traces during events without needing a bunch of infrastructure.

The downside is that if you wanted to see something that already happened, you can't. This is for looking at problems that are happening!

Currently this repo doesn't offer opentelemetry traces.

What this is for

This is suited for short-lived APIs — request-response work, RPC handlers, background jobs that complete in milliseconds to seconds. The cache only commits spans to its shared map when they close, and the protocol only streams closed spans. Anything still in flight is invisible to the console.

That makes it a poor fit for:

• Long-running background spans (event loops, daemons, things that stay open for hours) — they never reach the cache.
• Low-value logging that has no span structure — the console can stream events, but the UI is span-oriented.
• Distributed observability — This is a single-host debugging console, not a distributed tracing backend.

What it is good at: attaching to a running server, watching the live shape of recent requests, finding outliers, drilling into one trace's full event/span tree, and disabling again without leaving any undue overhead behind.

Getting started — integrating with your server

1. Add the dependencies

[dependencies]
tokio = "1"
tracing = "0"
tracing-cache = "0"
tracing-console-host = "0"

2. Stand up the cache + serve

use std::sync::Arc;
use tracing_cache::{ChancePredicate, LevelPredicate, SpanCache};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Predicate stack: LevelPredicate gates by min level; ChancePredicate
    // wraps it with a 0-100 % sampling rate.  Both are dynamic — the
    // console can flip them at runtime via the level switcher and the
    // `C` chance modal. You should probably always start this OFF.
    let level = LevelPredicate::with_filter(tracing::metadata::LevelFilter::OFF);
    let level_handle = level.handle();
    let predicate = ChancePredicate::new(level, 100.0);
    let chance_handle = predicate.handle();

    // 16k closed-span ring buffer. Sized for your worst-case burst rate
    // × the time you'd want to look back at it. If you overflow while
    // recording you'll just drop spans. Oh well.
    let (cache, driver) = SpanCache::with_predicate(16_384, predicate);
    let cache = Arc::new(cache);

    // Install as the global subscriber.  Spans your code emits anywhere
    // in the process from now on land in this cache when they close.
    tracing::subscriber::set_global_default(Arc::clone(&cache))?;
    tokio::spawn(driver.run());

    // Serve the cache to console clients.  Pick a port reachable from
    // wherever you'll be running the TUI.
    let addr: std::net::SocketAddr = "127.0.0.1:7890".parse()?;
    tokio::spawn(async move {
        if let Err(e) = tracing_console_host::serve(cache, level_handle, chance_handle, addr).await {
            eprintln!("tracing-console-host serve: {e}");
        }
    });

    // ... your server's normal entry point ...
    your_server::run().await
}

3. Instrument with regular tracing macros

You put your root spans carefully at the base of your APIs and the units of work you want to trace. For async code, use #[tracing::instrument] or future.instrument(span).await instead of let _g = span.enter(). An enter() guard sits on the thread until it goes out of scope, including across every .await in between; when the executor parks the task and runs something else on the same worker thread, that unrelated work ends up captured as a child of your span. That's really confusing.

use tracing::{event, Instrument, Level};

// Outer handler — root of one trace per request.  `parent = None`
// explicitly anchors it so it can't inherit whatever happened to
// be on the caller's stack.  `skip(req)` keeps the request body
// out of the span fields; we pull the bits we want via `fields`.
#[tracing::instrument(
    parent = None,
    skip(req),
    fields(method = %req.method, path = %req.path),
)]
async fn handle_request(req: Request) -> Response {
    event!(Level::INFO, "received");
    let body = parse_body(&req).await;       // each its own #[instrument]
    let result = run_query(&body).await;
    event!(Level::INFO, rows = result.rows, "done");
    build_response(result)
}

#[tracing::instrument(skip(body))]
async fn parse_body(body: &Body) -> Parsed { /* … */ }

If you'd rather build the span by hand, use .instrument:

let span = tracing::info_span!("request", method = %req.method);
do_work(&req).instrument(span).await

let _g = span.enter() is fine in synchronous code where no .await happens inside the guard's scope.

4. Connect from the TUI

cargo run --release -p tracing-console -- 127.0.0.1:7777

You'll land in the stacks view with nothing visible (level defaults to Off). Press Shift+I (or Shift+D, Shift+T) to ask the server to start recording at Info / Debug / Trace. Spans will start showing up as they close.

Keyboard cheat sheet

The UI highlights available options contextually. Here are some of the main controls:

Configuration

• Cache capacity (SpanCache::with_predicate(capacity, …)) — the max number of closed spans in the ring buffer. Older spans evict FIFO. Too small and your client will get gaps. Too large and you'll just be wasting memory for no reason while you're tracing.
• CacheConfig::pending_batch (default 8) — per-thread closed-span buffer before flushing. Lower = more responsive at low traffic, more sends at high traffic.
• CacheConfig::channel_capacity (default 65_536) — buffer between producer threads and the driver.
• Sampling rate (ChancePredicate) — live-tunable via Shift+C in the TUI, or programmatically via chance_handle.
• Cache level (LevelPredicate) — live-tunable via Shift+letter in the TUI, or programmatically via level_handle. The server starts at whatever you initialized; starting up with Off is recommended.

Examples

• tracing-console-host/examples/fs_listing_api.rs — a self-contained host that traces a directory walk.
• tracing-console-host/examples/synth_load.rs — a synthetic-load generator used for bench testing.

# terminal 1 — start the example "server"
cargo run -p tracing-console-host --example fs_listing_api -- /tmp

# terminal 2 — connect the TUI
cargo run -p tracing-console -- 127.0.0.1:7777

Limitations

Spans only become visible when they close. In-flight long-running spans don't show up until they end. When you enable tracing, you are doing it to a live system in who-knows what state. You'll get some orphaned spans at the beginning, and some incomplete spans at the end.

This tool is intended for quick live analysis, not rigorous archival. There is no persistence. All caching is done in bounded ring buffers in-memory, and lost on restart.