Deterministic Simulation Testing (DST)

Deterministic simulation testing (DST) runs NautilusTrader under a seed-controlled runtime so that timing-sensitive execution behavior is bitwise reproducible from a single integer. This page is the determinism contract: what the runtime guarantees under seed-controlled execution, the seams that implement those guarantees, the pre-commit hook that enforces them, and where they stop. Each claim names the source site behind it, so users and auditors can check the contract against the code.

:::note A downstream harness that depends on NautilusTrader's determinism consumes the version of this document at its pinned NautilusTrader commit. A change to this document is a contract change for those consumers and should be reviewed as one. :::

What DST is

DST is a testing technique for concurrent systems. A single seed fully determines an execution, including task scheduling, timer firings, and random values. Two runs with the same seed, binary, and configuration produce identical observable behavior. When a property fails, the seed is the reproduction: the same seed replays the failure every time.

Scheduling decisions in an async runtime come from ambient process state: task wake order, timer resolution, thread scheduling, hash seeds. None of that is controlled by the test harness, which is why a race that surfaces once in CI is usually hard to reproduce on demand. DST replaces those ambient sources with a seeded pseudorandom sequence, so the interleaving is a function of the seed.

FoundationDB applied the pattern to a production distributed database starting around 2009. In the Rust ecosystem, madsim intercepts tokio primitives to provide a deterministic scheduler.

DST targets the bugs that escape unit, integration, property, and acceptance testing: channel wakeup ordering, drain races at shutdown, startup sequencing, reconciliation ordering, and recovery-path correctness. Each involves interleavings that other test layers cannot exhaustively cover but a deterministic scheduler can explore systematically.

Goals

• Seed-reproducible execution for the in-scope portion of the NautilusTrader runtime.
• Honest scope. The contract lists what is covered and what is not. No silent fallbacks to real wall-clock time or unseeded RNG; conditions that weaken the guarantee are enumerated.
• Enforcement in source. A pre-commit hook fails commits that add banned patterns to the DST path, so the contract stays true without relying on reviewer attention.
• Minimum necessary instrumentation. The seams route time, task scheduling, and randomness through a deterministic source only where the contract requires it; everything else runs unchanged.

Approach

madsim determinizes only the tokio primitives that route through its aliased submodules (time, task, runtime, signal). Wall-clock reads, monotonic reads, RNG draws, hash iteration, and select! polling bypass tokio entirely and need their own seams. Layer 1 swaps the aliased submodules for madsim; Layer 2 supplies the seams.

Layer 1: runtime swap

Under the simulation Cargo feature on nautilus-common, four tokio submodules are routed through madsim when RUSTFLAGS="--cfg madsim" is set:

• time (timers, intervals, monotonic Instant).
• task (spawning and joining async tasks).
• runtime (the runtime builder and handle).
• signal (process signals such as ctrl_c). The re-export exists; call-site adoption is partial (see Signal handling).

These re-exports live in nautilus_common::live::dst. DST-path call sites for time, task, and runtime import from this module rather than directly from tokio, so toggling the feature switches the async runtime in one place for the primitives that are fully routed. Under normal builds, the re-exports resolve to real tokio. Under simulation + cfg(madsim), they resolve to madsim's deterministic counterparts.

Everything else that tokio provides (sync, io, select! as a macro, fs, net) uses real tokio unconditionally. Transitive crates (tokio-tungstenite, tokio-rustls, reqwest) are unaffected.

Layer 2: nondeterminism substitution

Nondeterminism outside the async runtime is redirected through explicit seams:

• Wall-clock reads go through nautilus_core::time::duration_since_unix_epoch. Under simulation this routes to madsim::time::TimeHandle::try_current(), preserving Unix-epoch semantics for order and fill timestamps. When called outside a madsim runtime (plain #[rstest] test bodies), it falls back to SystemTime::now(), which under cfg(madsim) is libc-intercepted to the same real syscall a normal build would use. Production paths under simulation always run inside a runtime, so they continue to receive virtual time.
• Monotonic reads go through nautilus_common::live::dst::time::Instant. The type resolves to tokio::time::Instant on normal builds, which keeps tokio::test(start_paused) tests working, and to madsim::time::Instant under simulation.
• Network-local monotonic reads go through nautilus_network::dst::time. The crate sits below nautilus-common in the dependency graph and exposes a local re-export module with the same semantics.
• Hash iteration order in the reconciliation manager and the order matching engine uses IndexMap and IndexSet rather than AHashMap and AHashSet. AHash randomizes its hasher per process; insertion-order iteration is needed where order drives downstream event publication or the sequence in which a seeded FillModel RNG is consumed.
• tokio::select! polling order uses the biased; modifier at every production site on the DST path. Unbiased select! polls branches in an order chosen by an unintercepted RNG.

Determinism contract

Under the conditions below, a run identified by (seed, binary hash, configuration hash) on the same platform produces bitwise-identical:

• Scheduling order of async tasks.
• Timer firings (virtual monotonic and virtual wall-clock).
• RNG output from madsim::rand.
• Delivery order on tokio::sync channels.

Required conditions

The contract holds only when all of the following are true:

• The simulation Cargo feature is active and RUSTFLAGS="--cfg madsim" is set. Both are required. The feature activates the deterministic runtime; the cfg flag activates madsim's libc-level clock_gettime and getrandom intercepts. One without the other silently falls back to real tokio and breaks determinism without an error.
• Every tokio::select! call site on the DST path uses the biased; modifier.
• Monotonic time reads route through the DST seam (either nautilus_common::live::dst::time or nautilus_network::dst::time), not std::time::Instant::now directly.
• Wall-clock time reads route through nautilus_core::time::duration_since_unix_epoch.
• Randomness routes through madsim::rand. rand::thread_rng, rand::rng(), fastrand, getrandom, and OsRng are not intercepted.
• Iteration-order-sensitive collections either use IndexMap or IndexSet, or sort at the point of use.
• tokio::task::LocalSet construction is cfg-gated out under simulation. madsim does not provide LocalSet; spawn_local works without it.
• tokio::task::spawn_blocking call sites are cfg-gated or removed. A blocking call escapes the deterministic scheduler.

Static enforcement

Static enforcement has two layers:

• Clippy policy in clippy.toml and [workspace.lints.clippy] blocks APIs that are invalid across the workspace DST contract: direct getrandom::{fill,u32,u64} calls and tokio::task::LocalSet.
• A pre-commit hook named check-dst-conventions enforces scoped, path-aware, and cfg-aware structural checks that Clippy cannot express cleanly.

The hook lives at .pre-commit-hooks/check_dst_conventions.sh and runs both in the standard pre‑commit suite and in CI. Rules 1 to 6 apply to the 17 in‑scope workspace crates; Rule 7 applies to the nine crates on the madsim build path. The hook fails the commit when a rule detects:

• Rule 1: raw std::time::Instant::now(), SystemTime::now(), or chrono::Utc::now() reads, including bare forms when the enclosing file imports the type from std::time, or from chrono for Utc.
• Rule 2: raw RNG usage (rand::thread_rng, rand::rng(), fastrand::, getrandom::, OsRng) or Uuid::new_v4() without cfg gating.
• Rule 3: tokio::select! blocks missing biased; within the first three lines.
• Rule 4: std::thread::spawn, std::thread::Builder::new, or tokio::task::spawn_blocking calls that lack a preceding #[cfg(test)], #[cfg(not(madsim))], or #[cfg(not(all(feature = "simulation", madsim)))] attribute.
• Rule 5: AHashMap or AHashSet in iteration-order-sensitive files on the DST path. Enforcement covers the two audited files, crates/live/src/execution/manager.rs and crates/execution/src/matching_engine/engine.rs; the full file set remains under audit.
• Rule 6: direct tokio::net::TcpStream::connect / tokio::net::TcpListener::bind reaches that bypass nautilus_network::net. The seam re-exports tokio::net types under normal builds and swaps to turmoil::net under the turmoil feature, so all TCP entry points share a single cfg-gated swap point.
• Rule 7: raw tokio::{time,task,runtime,signal} paths in production code on the madsim build path. Callers route these modules through nautilus_common::live::dst; the facade definition, process-wide real Tokio runtime, and test infrastructure are explicit exceptions.

The hook supports two exception forms:

• An inline // dst-ok marker on a specific line, typically accompanied by a short reason (for example, log-only wall-clock timing that does not affect state).
• A small file-level allowlist in the hook script itself for sites classified as leave-alone in the codebase audit (log timing in the cache module, log-record timestamping in the logging bridge and writer, progress reporting in the DeFi module).

Test files, files under tests/, python/, and ffi/ directories, and lines inside an inline #[cfg(test)] module are excluded because they are not part of the DST path.

In-scope crates

The hook applies to 17 workspace crates: the 16 crates in the transitive closure of nautilus-live (analysis, common, core, cryptography, data, execution, indicators, live, model, network, persistence, portfolio, risk, serialization, system, and trading), plus backtest.

Adapter crates and infrastructure crates (Redis, Postgres) are out of scope. Their DST suitability requires a separate audit before they enter the DST path.

Network seed soaks

Turmoil simulates the network under a seeded scheduler, so the nautilus-network transport tests can explore link and reconnect orderings that the madsim runtime swap does not reach. These tests run in two layers:

• Fixed-seed tests run in the nightly test suite. These cover connect, reconnect, partition, close during reconnect, close during backoff, and repeated server-drop scenarios with reproducible seeds.
• An ignored reconnect soak sweeps Turmoil seeds until stopped, or until NAUTILUS_TURMOIL_SOAK_COUNT seeds have run. Each seed runs the Tungstenite WebSocket backend first and the Sockudo backend second when transport-sockudo is enabled, so both backends see the same schedule search path.

Run the continuous soak with:

scripts/soak-network-turmoil.sh

Run a bounded soak with:

env NAUTILUS_TURMOIL_SOAK_COUNT=100 scripts/soak-network-turmoil.sh

NAUTILUS_TURMOIL_SOAK_START sets the first seed, which resumes a sweep where an earlier run stopped. Within each seed, Turmoil runs the nodes in random order and randomizes link latency between 1 ms and 25 ms, while the scenario drops the server repeatedly, cycles the client through reconnect states, and asserts exact application-message order. The soak does not enable Turmoil fail_rate: for TCP, that breaks links without a retransmit model, which would overstate the client delivery contract for an order-preservation test.

The Turmoil tests run against the simulated network and are not gated to Linux, so a seed sweep runs anywhere, including macOS. Several real localhost socket and WebSocket unit tests use target_os = "linux" for CI stability, so a macOS run leaves that host TCP coverage untouched. Treat the network test set as covered only after a run on Linux CI or a Linux workstation.

Implementation notes

Concrete changes the DST audit produced. Start here when checking whether a code path is on the DST path and how it routes.

Iteration-order seams

Production sites that hold IndexMap / IndexSet rather than AHashMap / AHashSet because the iteration order is observable on the DST path:

• Matching engine (crates/execution/src/matching_engine/engine.rs): ten fields (execution_bar_types, execution_bar_deltas, account_ids, cached_filled_qty, bid_consumption, ask_consumption, queue_ahead_orders, queue_ahead_total, queue_excess, queue_pending). Iterated removes use .shift_remove(). Closes #3914.
• Reconciliation manager (crates/live/src/execution/manager.rs): hook-enforced; the ReconciliationResult report maps (orders, fills) live in crates/execution/src/reconciliation/types.rs as IndexMap.
• Account trait (crates/model/src/accounts/): balances, balances_total, balances_free, balances_locked, starting_balances returns. Balance and margin storage fields on BaseAccount and MarginAccount are IndexMap; commissions and leverages remain AHashMap.
• Position events (crates/model/src/position.rs): Position::commissions, consumed via .values() in events/position/snapshot.rs.
• Portfolio aggregation (crates/portfolio/src/portfolio.rs): unrealized_pnls, realized_pnls, net_positions storage; accumulate_mark_values builds IndexMap<Currency, Decimal>.
• Data engine (crates/data/src/engine/): book_snapshot_counts, bar_aggregators, BookSnapshotInfos. Iterated removes use .shift_remove().
• Execution engine (crates/execution/src/engine/): ExecutionEngine.clients, plus the client_ids / venues accumulators in get_clients_for_orders().
• Backtest engine and exchange (crates/backtest/src/engine.rs, crates/backtest/src/exchange.rs): BacktestEngine.venues and SimulatedExchange.matching_engines preserve venue and instrument iteration order for settlement, expiration, liquidation, and seeded FillModel draws (#4480).
• Trading algorithm (crates/trading/src/algorithm/core.rs): strategy_event_handlers (drives ordered msgbus::unsubscribe_* fan-out).
• Analyzer (crates/analysis/src/analyzer.rs): account_balances, account_balances_starting.
• Cache API (crates/common/src/cache/mod.rs): get_orders_for_ids and get_positions_for_ids sort their Vec returns by client_order_id / position_id before returning. Storage stays on AHashSet (set semantics).
• Instrument store (crates/common/src/providers.rs): InstrumentStore.instruments, because the Betfair, Derive, and Polymarket adapters publish one DataEvent::Instrument per entry straight from get_all() / list_all(). Keeps the ahash hasher.
• Order emulator (crates/execution/src/order_emulator/emulator.rs): on_reset sorts the drained subscribed_quotes, subscribed_trades, and subscribed_strategies before the msgbus::unsubscribe_* fan-out. The quote and trade paths also advance the seeded UUID4::new draw sequence. Storage stays on AHashSet.
• WebSocket subscriptions (crates/network/src/websocket/subscription.rs): topics_from_map sorts its Vec return, which fixes the reconnect replay order behind all_topics(). Storage stays on DashMap with AHashSet values.

Remaining AHashMap / AHashSet sites in the original nautilus-live closure are lookup‑only, behind concurrent shared‑ownership wrappers (Arc<DashMap>, AtomicMap), or feed into commutative aggregation. backtest retains additional hash collections outside rule 5's two‑file enforcement scope, including pre‑run validation and result maps. Treat their iteration order as outside the static guarantee until each path is audited.

Time seams

Instant::now / SystemTime::now call sites that remain on the DST path are either inside #[cfg(test)], file-allowlisted in the hook, or carry an inline // dst-ok marker with a reason:

• crates/common/src/testing.rs: wait_until / wait_until_async timers.
• crates/execution/src/engine/mod.rs: init log timing in load_cache.
• crates/common/src/cache/mod.rs: timing in check_integrity and audit_own_order_books (file-allowlisted).
• crates/model/src/defi/reporting.rs: progress logging (file-allowlisted).
• crates/core/src/time.rs: seam definition site (file-allowlisted).

chrono::Utc::now is hook-banned in the in-scope crates. The remaining call sites are the logging bridge and writer, scoped out under Logging runs on real OS threads. crates/core/src/datetime.rs::is_within_last_24_hours routes through nautilus_core::time::nanos_since_unix_epoch() and compares in u64 nanos directly.

Randomness seams

Production RNG sites on the DST path:

• crates/core/src/uuid.rs::UUID4::new() routes through madsim::rand::thread_rng() when called inside a madsim runtime under simulation, falling back to rand::rng() outside one (and on normal builds). Production paths under simulation always run inside a runtime, so they consume seeded bytes; plain #[rstest] tests under cfg(madsim) use the host RNG. Reachable from order and event factories in nautilus-common and nautilus-risk.
• crates/execution/src/models/fill.rs::default_std_rng() routes the same way. Called from ProbabilisticFillState::new() when no seed is provided. With a seed, StdRng::seed_from_u64 is deterministic by construction.
• crates/execution/src/matching_engine/ids_generator.rs uses nautilus_core::UUID4::new() in the position and venue order id generators for the use_random_ids path. The default ID scheme ({venue}-{raw_id}-{count}) is deterministic without it.

One site carries a marker: jitter sampling for reconnect backoff in crates/network/src/backoff.rs, marked // dst-ok as transport layer.

Tokio submodule split

madsim aliases time, task, runtime, and signal. The other tokio submodules (sync, io, select!, fs, net) stay on real tokio under simulation. Extending the swap further would require rebuilding tokio-tungstenite, tokio-rustls, and reqwest against shimmed tokio::net::TcpStream, which the audit ruled out as too invasive.

In-scope sites that touch real tokio::net / tokio::io directly:

• crates/network/src/net.rs re-exports tokio::net::{TcpListener, TcpStream}.
• crates/network/src/socket/client.rs uses tokio::io::{AsyncReadExt, AsyncWriteExt}.
• crates/network/src/tls.rs uses tokio::io::{AsyncRead, AsyncWrite}.
• crates/network/src/socket/types.rs aliases MaybeTlsStream<TcpStream> split halves via tokio::io::{ReadHalf, WriteHalf}; the TCP type itself comes through the crate::net seam.

These run on real sockets even under simulation. Channel delivery order on tokio::sync stays deterministic because the sender and receiver tasks are scheduled by the madsim executor even though the channel implementation is real.

Raw thread escape rules

Rule 4 of the hook bans raw thread spawning outside three escape cases:

• #[cfg(test)] test modules.
• #[cfg(not(madsim))] or #[cfg(not(all(feature = "simulation", madsim)))] production sites (e.g. the logging writer thread).
• An inline // dst-ok marker.

tokio::task::LocalSet and tokio::task::spawn_blocking are not supported under madsim. The codebase audit found no production sites for either inside the in-scope crates; new sites must carry a cfg gate or // dst-ok marker.

Logging tests under simulation

The logging writer thread is cfg-gated out under simulation; under cfg(madsim) log events are dropped. Tests that init the file-logging writer would either hang or assert against an empty log file, so the affected submodules are gated out at the module boundary:

• crates/common/src/logging/logger.rs::tests::serial_tests.
• crates/common/src/logging/macros.rs::tests.

logger.rs::tests::sim_tests::test_init_under_madsim_skips_writer_thread_and_forces_bypass runs under simulation and pins the gated behavior.

Scope boundaries

The contract is deliberately narrow. The following weakenings are explicit, not oversights.

Python is not in DST scope

DST runs under a native Rust test harness. No Python interpreter starts during a DST run. The PyO3 bindings under crates/*/src/python/, the ffi/ directories, and the Python packages under nautilus_trader/ are excluded from the contract as a policy, not as a weakness. Any code reachable only from Python call paths is out of scope; any Rust path reachable from the native DST harness must satisfy the contract even if the same type is also exported to Python.

The check-dst-conventions hook encodes this policy by skipping /python/ and /ffi/ paths in the in-scope crates. Clock, RNG, and threading call sites behind those paths do not apply to the contract.

The primary objective of DST is reliability of the Rust engine itself: the order lifecycle, reconciliation, matching, risk, and execution state machines. Deterministic replay of user strategies is a secondary goal, reachable only for strategies authored in Rust or driven through a Rust-native test harness. A Python strategy that calls time.time(), issues arbitrary network requests, or relies on thread scheduling can vary its command stream between runs; the Rust core processes the varying stream deterministically, but end-to-end replay from a Python entry point is not guaranteed.

Platform-scoped

madsim's libc overrides for clock_gettime and getrandom are platform-specific. Cross-platform bitwise reproducibility is not claimed. A seed that reproduces a failure on Linux x86_64 may not reproduce on macOS aarch64.

Non-aliased dependencies escape silently

Any dependency that reaches the OS through a non-aliased path (direct libc calls, std::net bypass, crates using fastrand or OsRng) escapes the simulator without raising an error. The in-scope crates have been audited; adapter crates and infrastructure crates require their own audits before entering the DST path.

Transport-layer I/O is not simulated

tokio-tungstenite, tokio-rustls, reqwest, redis, and sqlx use real tokio internally. Under simulation, WebSocket and HTTP I/O run on real networking. This is intentional: the initial target is order lifecycle determinism, not transport fault injection. Transport-layer determinism would require per-crate madsim shims that do not exist.

Test modules that drive real localhost sockets (crates/network/src/socket/client.rs::tests, ::rust_tests; crates/network/src/websocket/client.rs::tests, ::rust_tests; crates/network/tests/websocket_proxy.rs) are cfg-gated out under all(feature = "simulation", madsim) because their production code paths reach dst::time::* (madsim time primitives), which panic when called from a #[tokio::test] runtime. The retry test modules (crates/network/src/retry.rs::tests, ::proptest_tests) run under simulation: each test attribute is cfg_attr-swapped between #[tokio::test(start_paused = true)] and #[madsim::test], time reads and sleeps route through crate::dst::time, and explicit virtual-time advances go through a cfg-gated advance_clock function so the same body covers both runtimes.

Signal handling

nautilus_common::live::dst::signal exposes routed ctrl_c and terminate re-exports. The crates/live/src/node/mod.rs run loop routes through them, so node shutdown driven by ctrl_c is injectable from test code under cfg(madsim) via madsim::runtime::Handle::send_ctrl_c. Adapter-bin entry points still call tokio::signal::ctrl_c directly and remain scoped out.

Logging runs on real OS threads

The logging subsystem spawns a writer thread via std::thread::Builder and uses std::sync::mpsc. Under simulation, the thread is not spawned and log events are dropped. Log output is outside the determinism contract: the writer only writes, never reads or mutates simulation state.

Adapters

Adapter crates are out of scope. They carry their own direct clock, RNG, and transport-layer call sites (chrono::Utc::now, SystemTime::now, raw transport clients), varying by adapter. An adapter that enters the DST path must be audited for those call sites before the contract covers its behavior.

Relationship to other testing layers

DST complements existing testing; it does not replace any of it.

DST's unique value is in the intersection of async concurrency and state-machine correctness. Bugs such as "a message at shutdown is dropped under a specific wakeup ordering" or "a reconciliation event is lost when iteration order reverses" are the target class. For anything else, the pre-existing test layers are the right tool.

Status

• Layer 1 (runtime swap) is implemented. nautilus_common::live::dst exposes routed re-exports for time, task, runtime, and signal. Production call sites for time, task, and runtime route through the seam; signal call-site adoption is partial (see Signal handling).
• Layer 2 (nondeterminism substitution) is implemented across the 17 in‑scope crates. Seams exist for wall-clock time, monotonic time, randomness, and iteration order. Implementation notes enumerates the audit closures and the remaining allowed call sites.
• Static enforcement via check-dst-conventions is active in pre-commit and CI. The hook covers the load-bearing conditions; the // dst-ok marker convention permits per-line exceptions when justified.
• End-to-end runtime verification (a same-seed diff over an in-scope code path) is out of scope for this repository. The hook enforces the structural conditions; the claim that a seed reproduces identical observable behavior across runs is plausible from the seam design but is not verified by a regression gate.

Simulation smoke gate

The dst workflow (.github/workflows/dst.yml) invokes make cargo-test-sim, which builds nautilus-common, nautilus-core, nautilus-network, nautilus-execution, and nautilus-live with --features simulation under cfg(madsim), then runs the sim-compatible legs below. Each leg runs with its own crate's --features simulation and uses #[madsim::test] where applicable, so the explicit cfg branches and virtual time are both validated. nautilus-common and nautilus-execution consume nautilus-model types, so each also runs a second leg with --features "simulation,high-precision", exercising the seam-routed code paths under both fixed-point widths (QuantityRaw and PriceRaw as u64 and as u128).

The gate covers:

• All simulation-compatible nautilus-common tests. This leg compiles with nautilus-core/simulation propagated, so the explicit wall_clock_now cfg branch is selected for every test in the suite. Plain #[rstest] tests run outside a madsim runtime and route through the seam's SystemTime::now() fallback, the same path madsim's libc shim takes outside a runtime. The LiveClock test module is cfg-gated out because its plain #[rstest] cases start LiveTimer tasks without a madsim runtime, and most also block on wall-clock progress. live::dst::tests::test_dst_wall_clock_advances_with_virtual_time uses #[madsim::test] and asserts that nanos_since_unix_epoch advances with madsim::time::sleep, so virtual wall-clock behavior is validated end-to-end on this leg.
• The nautilus-live startup reconciliation timeout regression. This runs under a madsim runtime and verifies that a pending mass-status request reaches the configured timeout, reports the expected error, and cleans up the node instead of entering a real Tokio timer.
• All of nautilus-network, with transport-bound test modules gated out at the source. Includes the seam pinning tests for sleep, timeout, and the rate limiter under virtual time, plus the retry suites that exercise backoff timing.
• All of nautilus-execution. These are plain #[rstest] cases, so they compile and exercise the cfg-gated branches in the matching engine, fill model, and execution-engine state machines without entering a madsim runtime; default_std_rng() takes its host-RNG fallback there.
• The cross-crate seam pinning tests in nautilus-core (wall_clock_now virtual time).

Deterministic-scheduler coverage comes from the #[madsim::test] cases in nautilus-common, nautilus-core, nautilus-network, and nautilus-live. The gate as a whole catches drift in the cfg-gated DST seams; it does not verify determinism end-to-end.

Further reading

• .pre-commit-hooks/check_dst_conventions.sh defines the seven enforcement rules in full and documents the // dst-ok marker convention.
• FoundationDB testing philosophy.
• TigerBeetle simulation testing blog posts.
• madsim repository, the deterministic runtime.
• Turmoil repository, the deterministic network simulator.