History
Origin
Chaosblade grew out of Alibaba's internal fault-testing and drill tooling, codenamed MonkeyKing. Alibaba's own account describes roughly ten years of fault-injection practice behind it, starting from validating the dependency problems of microservices and widening to steady-state validation in cloud and cloud-native environments. The project was open-sourced in 2019, initially as two repositories: chaosblade, the Go CLI plus basic resource and container executors, and chaosblade-exec-jvm, the executor for the JVM.
The stated motivation was that chaos tools of the time had scattered scenarios, were hard to deploy, lacked a standard experiment model, and were difficult to extend. Chaosblade's answer was to standardise the experiment model so scenarios could be defined, shared, and extended uniformly.
Timeline
| Year | Milestone |
|---|---|
| 2019 | Open-sourced by Alibaba as chaosblade (Go CLI) plus chaosblade-exec-jvm |
| 2020 | v1.0.0-GA: Linux, Windows, Docker, and Kubernetes support; Java, Golang, JavaScript, C++ coverage; the chaosblade-box platform added |
| 2021 | Accepted into the CNCF Sandbox on 2021-04-28 |
| 2022 | New version of the chaosblade-box chaos engineering platform released |
| 2026 | Python blade-ai/ agent layer added; mainline release line at v1.8.0, AI layer at blade-ai-v0.5.1 |
How it evolved
The project moved from a single tool to a platform. Alibaba describes adding chaosblade-box to manage chaos across multiple clusters, environments, and languages, and reports that at the v1.0.0-GA mark the tool covered 200+ scenarios with 3000+ parameters. The architecture stayed split: a thin CLI plus a family of sibling executor repositories (chaosblade-exec-os, chaosblade-exec-jvm, chaosblade-exec-cplus, and so on) that the build clones and packages, so new fault domains arrive as new executors rather than CLI rewrites.
Two visible changes appear in the source at this commit. Server command mode was disabled on 2023-12-30, noted in the CLI init at cli/cmd/cmd.go:58, so the CLI no longer registers a server subcommand even though the README still mentions it. And a Python AI agent layer was added under blade-ai/, described as orchestration that takes a plain-language failure description and runs intent understanding, safety review, injection, verification, recovery, and reporting.
Where it stands now
Chaosblade remains a CNCF Sandbox project; the same-category projects Chaos Mesh and LitmusChaos are one step further at Incubating. The 2025 survey paper (arXiv 2505.13654) positions ChaosBlade as Sandbox with active development across many releases. LFX Insights reports a single-contributor concentration risk: in a recent quarter one contributor accounted for more than half of activity. The mainline release line is at v1.8.0, with the newer AI layer versioned separately as blade-ai-v0.5.1 (2026-06-23). The Go module targets go 1.25 (go.mod:17).