Build the rollback mechanism, and test it
ADR 0059's recovery path: the pieces that run when the host will not start. internal/upgrade -- two facts, neither of them the host judging its health. Whether the executable this process started from has been replaced on disk, and which version last completed a reconcile. The first design was wrong and the tests caught it, not review. It asked /proc/self/exe whether it was marked deleted. That is Linux procfs behaviour rather than a fact about files, and it catches only unlink -- a binary swapped by rename onto the same path reads as untouched, which is exactly what a package manager does. Now the identity is captured at start and compared later: no procfs, and neither case missed. known-good is one bare line. The reader is a shell script on a machine where the host is failing to start, so it must not need a parser to be present and working. Written only after a clean apply, which is the whole claim -- not health, because a disconnected node is ordinary and a failing resource is the machine's problem rather than the binary's. packaging/ -- the unit, the rollback unit, and the rollback script. The script shares no code with the host and calls none of it: a binary that cannot start cannot be its own recovery. POSIX sh, nothing that has to be installed. The unit carries Restart=always with a comment saying why on-failure would break every upgrade. Both are tested and both sets of tests were confirmed to bite. Injecting five faults broke exactly the intended tests -- except one, and chasing why it did not found a placebo assertion I had written: `check "exits zero" ... "0" "0"` compares a literal to itself and can never fail. Replaced with the real exit code, after which the injection bites. Also caught: an injection that produced a build failure rather than a test failure, which my grep read as "no failure". Re-run so it compiled, and the test did bite. The script test runs in `make check`, so it is a gate rather than something that was run once. Verified against the real binary: known-good is written beside the store after a clean apply and is NOT written after a failed one.
This commit is contained in:
@@ -0,0 +1,138 @@
|
||||
// Package upgrade is how the host survives replacing itself.
|
||||
//
|
||||
// novox/hq ADR 0057 and ADR 0059. Two facts, and neither is the host judging its own health:
|
||||
//
|
||||
// - whether the executable this process started from has been replaced on disk, which is how
|
||||
// it knows to stand aside for a new one;
|
||||
// - which version last got as far as a completed reconcile, which is what a rollback outside
|
||||
// this binary reads when this binary will not start.
|
||||
//
|
||||
// The second is written for a reader that is not the host. A binary that cannot start cannot be
|
||||
// its own recovery, so what it leaves behind has to be plain enough for a shell script.
|
||||
package upgrade
|
||||
|
||||
import (
|
||||
"errors"
|
||||
"fmt"
|
||||
"os"
|
||||
"path/filepath"
|
||||
"strings"
|
||||
)
|
||||
|
||||
// KnownGoodName is the file a rollback script reads. Next to the store, because it is node
|
||||
// state of exactly the same kind.
|
||||
const KnownGoodName = "known-good"
|
||||
|
||||
// Self is the executable this process started from, remembered.
|
||||
//
|
||||
// Identity is taken once, at start, and compared later. The obvious alternative — asking
|
||||
// /proc/self/exe whether it is marked deleted — was tried and is worse in two ways: it is Linux
|
||||
// procfs behaviour rather than a fact about files, and it catches only *unlink*, so a binary
|
||||
// swapped by rename onto the same path reads as untouched. Remembering what we started from
|
||||
// needs no special filesystem and misses neither case.
|
||||
type Self struct {
|
||||
path string
|
||||
info os.FileInfo
|
||||
}
|
||||
|
||||
// Current captures the running executable's identity.
|
||||
//
|
||||
// path is what os.Executable() returned; a test passes one it can manipulate, because the
|
||||
// boundary being tested is the filesystem and a fake would assert that the fake behaves as
|
||||
// expected (novox/hq ADR 0034).
|
||||
func Current(path string) (Self, error) {
|
||||
info, err := os.Stat(path)
|
||||
if err != nil {
|
||||
return Self{}, fmt.Errorf(
|
||||
"cannot stat %s, so this host cannot tell whether it is later replaced: %w", path, err)
|
||||
}
|
||||
return Self{path: path, info: info}, nil
|
||||
}
|
||||
|
||||
// Path is where the executable was when this process started.
|
||||
func (s Self) Path() string { return s.path }
|
||||
|
||||
// Replaced reports whether a different file is at that path now, or none.
|
||||
//
|
||||
// Never a silent false: a host that cannot read its own image says so rather than assuming it is
|
||||
// current, which is the shape of every fault this repository catalogues.
|
||||
func (s Self) Replaced() (bool, error) {
|
||||
if s.info == nil {
|
||||
return false, errors.New("this host never captured its own identity, so it cannot tell " +
|
||||
"whether it has been replaced")
|
||||
}
|
||||
|
||||
now, err := os.Stat(s.path)
|
||||
if errors.Is(err, os.ErrNotExist) {
|
||||
// Removed rather than upgraded. Still not what is running, and saying "unchanged"
|
||||
// would leave the host claiming a version that is no longer installed.
|
||||
return true, nil
|
||||
}
|
||||
if err != nil {
|
||||
return false, err
|
||||
}
|
||||
|
||||
return !os.SameFile(s.info, now), nil
|
||||
}
|
||||
|
||||
// KnownGoodPath is where the marker lives, given where the store lives.
|
||||
func KnownGoodPath(statePath string) string {
|
||||
return filepath.Join(filepath.Dir(statePath), KnownGoodName)
|
||||
}
|
||||
|
||||
// RecordKnownGood marks a version as one that started and completed a reconcile.
|
||||
//
|
||||
// Written atomically and as one bare line. The reader is a shell script running on a machine
|
||||
// where the host is failing to start, so the format is the least it can be: no JSON, no
|
||||
// escaping, nothing that needs a parser to be present and working.
|
||||
func RecordKnownGood(path, version string) error {
|
||||
if strings.TrimSpace(version) == "" {
|
||||
return errors.New("refusing to record an empty version as known-good: a rollback " +
|
||||
"reading it would install nothing and report success")
|
||||
}
|
||||
if strings.ContainsAny(version, "\n\r") {
|
||||
return fmt.Errorf("refusing to record %q as known-good: it must be one line", version)
|
||||
}
|
||||
if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
|
||||
return err
|
||||
}
|
||||
|
||||
tmp, err := os.CreateTemp(filepath.Dir(path), ".known-good-*")
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
defer os.Remove(tmp.Name())
|
||||
|
||||
if _, err := fmt.Fprintln(tmp, version); err != nil {
|
||||
tmp.Close()
|
||||
return err
|
||||
}
|
||||
if err := tmp.Sync(); err != nil {
|
||||
tmp.Close()
|
||||
return err
|
||||
}
|
||||
if err := tmp.Close(); err != nil {
|
||||
return err
|
||||
}
|
||||
if err := os.Chmod(tmp.Name(), 0o644); err != nil {
|
||||
return err
|
||||
}
|
||||
return os.Rename(tmp.Name(), path)
|
||||
}
|
||||
|
||||
// ReadKnownGood returns the recorded version, or "" if there has never been one.
|
||||
//
|
||||
// Absence is not an error. A machine whose host has never completed a reconcile has no version
|
||||
// to go back to, and that is a real state rather than a fault: the node was never working, so
|
||||
// the failure belongs to the installation and not to an upgrade. A rollback that guessed here
|
||||
// would become a second fault.
|
||||
func ReadKnownGood(path string) (string, error) {
|
||||
raw, err := os.ReadFile(path)
|
||||
if errors.Is(err, os.ErrNotExist) {
|
||||
return "", nil
|
||||
}
|
||||
if err != nil {
|
||||
return "", err
|
||||
}
|
||||
return strings.TrimSpace(string(raw)), nil
|
||||
}
|
||||
@@ -0,0 +1,193 @@
|
||||
package upgrade
|
||||
|
||||
import (
|
||||
"os"
|
||||
"path/filepath"
|
||||
"strings"
|
||||
"testing"
|
||||
)
|
||||
|
||||
// started puts a binary on disk and captures it the way the host does at start.
|
||||
//
|
||||
// Against the real filesystem rather than a fake one. What is being tested is how the operating
|
||||
// system behaves when a file is replaced under a running process, and a fake would assert that
|
||||
// the fake behaves as expected (novox/hq ADR 0034).
|
||||
func started(t *testing.T) (Self, string) {
|
||||
t.Helper()
|
||||
binary := filepath.Join(t.TempDir(), "mesh-host")
|
||||
if err := os.WriteFile(binary, []byte("version one"), 0o755); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
self, err := Current(binary)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
return self, binary
|
||||
}
|
||||
|
||||
func TestAnUntouchedBinaryIsNotReplaced(t *testing.T) {
|
||||
// The case that runs every ten minutes forever. A false positive here is a node that exits
|
||||
// and restarts on every reconcile — a restart loop dressed as an upgrade.
|
||||
self, _ := started(t)
|
||||
|
||||
replaced, err := self.Replaced()
|
||||
if err != nil {
|
||||
t.Fatalf("could not tell: %v", err)
|
||||
}
|
||||
if replaced {
|
||||
t.Error("an untouched binary was reported as replaced; this host would restart forever")
|
||||
}
|
||||
}
|
||||
|
||||
func TestRewritingTheSameFileIsNotAReplacement(t *testing.T) {
|
||||
// Touching content in place keeps the inode, and a package manager does not install this
|
||||
// way — but something else on the machine might. The claim is about identity, not content.
|
||||
self, binary := started(t)
|
||||
|
||||
f, err := os.OpenFile(binary, os.O_WRONLY, 0o755)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if _, err := f.WriteString("same inode, new bytes"); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
f.Close()
|
||||
|
||||
replaced, err := self.Replaced()
|
||||
if err != nil {
|
||||
t.Fatalf("could not tell: %v", err)
|
||||
}
|
||||
if replaced {
|
||||
t.Error("writing through the same inode was reported as a replacement")
|
||||
}
|
||||
}
|
||||
|
||||
func TestInstallingOverTheBinaryIsAReplacement(t *testing.T) {
|
||||
// What a package manager actually does: write a new file and rename it over the old one.
|
||||
// The running process keeps the old inode; the path now holds a different file.
|
||||
self, binary := started(t)
|
||||
|
||||
next := binary + ".new"
|
||||
if err := os.WriteFile(next, []byte("version two"), 0o755); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if err := os.Rename(next, binary); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
replaced, err := self.Replaced()
|
||||
if err != nil {
|
||||
t.Fatalf("could not tell: %v", err)
|
||||
}
|
||||
if !replaced {
|
||||
t.Error("a binary replaced by rename was not noticed; this host would keep running the " +
|
||||
"old version and report the new one")
|
||||
}
|
||||
}
|
||||
|
||||
func TestRemovingTheBinaryIsAReplacement(t *testing.T) {
|
||||
// A package removed rather than upgraded. Nothing is at the path, and the honest answer is
|
||||
// still "not what I am running" — reporting unchanged would leave the host claiming a
|
||||
// version that is no longer installed.
|
||||
self, binary := started(t)
|
||||
if err := os.Remove(binary); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
replaced, err := self.Replaced()
|
||||
if err != nil {
|
||||
t.Fatalf("could not tell: %v", err)
|
||||
}
|
||||
if !replaced {
|
||||
t.Error("a removed binary was reported as unchanged")
|
||||
}
|
||||
}
|
||||
|
||||
func TestNotBeingAbleToTellIsAnError(t *testing.T) {
|
||||
// Never a silent false. A host that cannot read its own image must say so rather than
|
||||
// assume it is current, which is the shape of every fault this repository catalogues.
|
||||
if _, err := Current(filepath.Join(t.TempDir(), "no-such-binary")); err == nil {
|
||||
t.Fatal("capturing a nonexistent executable returned an identity instead of an error")
|
||||
}
|
||||
// And a Self that was never captured must refuse rather than answer.
|
||||
if _, err := (Self{}).Replaced(); err == nil {
|
||||
t.Fatal("an uncaptured Self answered instead of refusing")
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnownGoodRoundTrips(t *testing.T) {
|
||||
path := KnownGoodPath(filepath.Join(t.TempDir(), "state.json"))
|
||||
|
||||
if err := RecordKnownGood(path, "1.4.2"); err != nil {
|
||||
t.Fatalf("could not record: %v", err)
|
||||
}
|
||||
got, err := ReadKnownGood(path)
|
||||
if err != nil {
|
||||
t.Fatalf("could not read back: %v", err)
|
||||
}
|
||||
if got != "1.4.2" {
|
||||
t.Errorf("recorded 1.4.2 and read back %q", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestKnownGoodIsOneBareLine(t *testing.T) {
|
||||
// The reader is a shell script on a machine where the host is failing to start. It must not
|
||||
// need a JSON parser, and it must not need to strip anything but a newline.
|
||||
path := KnownGoodPath(filepath.Join(t.TempDir(), "state.json"))
|
||||
if err := RecordKnownGood(path, "1.4.2"); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
|
||||
raw, err := os.ReadFile(path)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if string(raw) != "1.4.2\n" {
|
||||
t.Errorf("known-good is %q; a rollback script reads this with `cat`, so it is one bare "+
|
||||
"line and nothing else", string(raw))
|
||||
}
|
||||
if strings.ContainsAny(string(raw), "{}\"") {
|
||||
t.Error("known-good contains structure; it must be readable without a parser")
|
||||
}
|
||||
}
|
||||
|
||||
func TestNeverHavingBeenGoodIsNotAnError(t *testing.T) {
|
||||
// A machine whose host has never completed a reconcile has nothing to go back to. That is a
|
||||
// real state — the node was never working — and a rollback must be able to tell it apart
|
||||
// from a read failure, because guessing a version is how recovery becomes a second fault.
|
||||
path := KnownGoodPath(filepath.Join(t.TempDir(), "state.json"))
|
||||
|
||||
got, err := ReadKnownGood(path)
|
||||
if err != nil {
|
||||
t.Fatalf("absence was reported as a failure: %v", err)
|
||||
}
|
||||
if got != "" {
|
||||
t.Errorf("expected no known-good version, got %q", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestAnEmptyVersionIsRefused(t *testing.T) {
|
||||
// An empty known-good would make the rollback script install nothing and report success —
|
||||
// the exact failure the rollback exists to prevent, relocated into the rollback.
|
||||
path := KnownGoodPath(filepath.Join(t.TempDir(), "state.json"))
|
||||
if err := RecordKnownGood(path, ""); err == nil {
|
||||
t.Fatal("an empty version was accepted as known-good")
|
||||
}
|
||||
}
|
||||
|
||||
func TestRecordingAgainReplacesRatherThanAppends(t *testing.T) {
|
||||
path := KnownGoodPath(filepath.Join(t.TempDir(), "state.json"))
|
||||
for _, v := range []string{"1.4.2", "1.4.3", "1.5.0"} {
|
||||
if err := RecordKnownGood(path, v); err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
}
|
||||
|
||||
got, err := ReadKnownGood(path)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if got != "1.5.0" {
|
||||
t.Errorf("after three recordings the file says %q; it holds the last one, not a history", got)
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user