Why it exists
Every Kotlin service starts with the same day of work before it does anything. keel is that day, done once: clone, rename, ./gradlew run, and there is a service — one route, one table, three probes, a /version, two targets both runnable, one image, every check green on day one.
It is konekt with the domain removed: the wiring a new service needs, and nothing the last one happened to have. It is not a library and publishes nothing, and it is not a framework — kore owns the lifecycle, sborka owns the build, and keel only shows them wired. The test for every file in it: did konekt or katcher need this? If not, it is not in keel.
What is in it
:server |
one KMP module, jvm() + linuxX64 (+ linuxArm64 behind a property) |
| one route | GET/POST /items, JSON via kotlinx.serialization |
| one store | ItemStore over sqlx4k-sqlite — the Rust driver on native, sqlite-jdbc on the JVM, one implementation |
| kore | probes, /version, and runUntilSignal with announce → drain → release |
| sborka | allocator settings, ktlint, a size budget, the staged binary path |
| zavarnik | a Leyden AOT cache for the JVM distribution, verified on check |
| the image | distroless/cc-debian13 by digest, the binary and nothing beside it, built with no Docker daemon |
docs/ |
a docs-bootstrap tree, passing make check on day one |
23 tests run on each of jvm and linuxX64 from one source, both halves pass kore's shutdown oracle, and the two binaries produce no diff after a normaliser written before the first run.
The numbers
Each began as a threshold declared before the first commit.
| target | measured | |
|---|---|---|
clone → /health/ready on both targets, cold machine, toolchain fetch included |
under 1 h | 3 min 48 s, image build not included |
image, distroless/cc |
under 25 MB | 13 972 497 bytes |
| Kotlin, main sources | under 500 lines | 214 |
| Gradle build logic | under 100 lines | 87 code lines |
the linuxX64 release binary |
— | 9 228 056 bytes, declaring 7 shared libraries |
docker images reports 55.4 MB for the same image, counting every platform of the base manifest; a budget checked with the wrong command fails at random, so the method is named next to the number. Going over a budget is the signal that something belongs in sborka or kore instead, and it is filed there rather than fixed here.
Starting a service from it
scripts/rename.sh relay # or: scripts/rename.sh webhook-relay com.example.relay
It rewrites every spelling of the name in the code — package, configuration prefix, types, binary, Docker paths, the workflow — and replaces the documentation rather than rewriting it, since keel's describes keel. Then it fails if git grep -il keel still finds anything.
Limitations
- kore, sborka and razves are not on Maven Central yet, so a clone needs the portfolio's Reposilite in
settings.gradle.kts, filtered toio.github.youndie. - A static image is not shipped, and not planned: about 4.5 MB of saving on an image already 44 % under budget, against a recipe that pins compiler properties a Kotlin bump can change.