Skip to content

Examples Overview

Tom Barber edited this page Oct 25, 2026 · 1 revision

Examples overview

The repository ships five worked examples under examples/ and docs/. They are real programs and real experiments, not toy demos, and they are honest about what they did and did not show. This page helps you pick one.

Each example has its own page with commands and expected output. Where a step needs Linux x86-64 and Hermit, the page says so and marks the output as copied from the repository's own documents rather than run by us.

At a glance

Example The bug Language What you learn How long What you need
Example-Dapr-orders A message consumer applies an order twice after a restart, because it remembers what it has seen in memory Python (app), Rust (simulator) How a simulated Dapr sidecar injects duplicates, reordering and crashes from a seed; how a fix is judged A couple of seconds per run Python 3, a Rust toolchain; runs on a Mac. Exploring seeds under Hermit needs Linux x86-64
Example-Language-fixtures A planted lost update (read, spin, write) C, C++, Go, Rust, Python, Java, .NET The same program per language; how each SDK is used; what the status of each language means Seconds per language; the Python self-check took nearly six minutes on our Mac The language's toolchain; runs natively on a Mac. Under Hermit: Linux x86-64
Example-Go-SQL-cancel-race A known database/sql race between cancelling a query and Rows.Scan (CVE-2025-47907, fixed in Go 1.24.6) Go Building one program three ways (buggy, fixed, patched-only-the-fix), an assisted control, how to read counts Minutes: about 8 seconds to fetch the toolchains and about 75 seconds to build the patched one on our Mac; the program itself runs in a fraction of a second Go; macOS arm64 for the helper scripts. Hermit part: Linux x86-64
Example-glibc-condlock A known glibc condition-variable lost wakeup (bug 25847) C How to design a property that does not fire falsely, plant faults to test the detector, and read an honest "we did not get a result" The native build and tests take seconds; the pinned-glibc experiment is long and needs a dedicated machine A C compiler for the native part. The full experiment: Linux x86-64, Hermit, a source build of glibc
Example-SQLite-WAL-reset A documented SQLite WAL data race (versions 3.7.0 to 3.51.2) C A research log with a clear "assisted, not unassisted" outcome About 20 seconds to fetch and build; the Hermit runs are long curl, a C compiler, python3. Hermit part: Linux x86-64

"How long" figures are what we measured on one Apple silicon Mac, or what the example's own documents say. Yours will differ.

Which one first

What these examples do and do not show

Read these before you quote anything from them.

  • The bug is known in advance. Each example is built around a published defect. Finding a bug you already know about is not the same as finding one you do not.
  • The workloads are targeted. They were written to put a specific race in reach. They are not generic tests of the library.
  • Some controls are assisted. An "assisted" variant widens the race window on purpose, to check the detector works. A result from one says the property and the detector work, and says nothing about how often the unassisted race is reached.
  • Counts are counts of runs in one setup. They are not rates of defects, and not claims about other machines, versions or tools. The Hermit-based numbers were measured on one bare-metal Linux machine with one pinned Hermit build.

More on how to read these claims: Limits-and-Honesty.

The Linux caveat, again

Anything described as "under Hermit" needs Linux x86-64 with hardware performance counters (Installation). On a Mac you can build and run the programs natively, run their unit tests, read their reports, and look at the SDK output, but you cannot run them as process targets. Each example page separates the steps that ran on a Mac from the ones that did not.

Clone this wiki locally