You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
feint shapes records what a real cloud returns and checks the emulator
against it. The field tree only — paths and JSON types, no values and no
identifiers — which is what makes it committable where a transcript is not: a
transcript describes somebody's account, a shape describes an API. Recording
needs a real account and stays a human's job; feint shapes --check compares
offline, with no credential, and reports how much it compared so a green
cannot be read as "nothing is wrong" when it means "nothing was checked".
internal/upstream, one place that knows how to talk to a real cloud —
signing, pacing and retrying for the three providers, answering with the same trace.Exchange the proxy writes. Seven copies of the Outscale signature had
accumulated in throwaway scripts, and the difference between two of them cost
an hour: one signed the request path, one signed /, and only the cloud could
tell them apart. Each signature now comes from its provider's own source.
Exoscale serves a lifecycle: instance start, stop, reboot, scale, resize
and delete, security groups and their rules, anti-affinity groups, elastic IPs
and their attachment — 16 to 46 operations served, 108 to 75 untriaged. Every
lifecycle path goes through Binding.Serialise, with a concurrency test under -race.
131 of 175 routes are proven by a real client, up from 109 of 145.
Fixed
data.outscale_images segfaulted the Terraform provider. Reported by
Vincent Dislaire, who traced it to the provider's own sources: three fields
read without a nil guard in a loop where every neighbour survives one, and the
catalogue published none of them. He measured the fix rather than guessing it,
injecting an empty BlockDeviceMappings through a proxy and watching the crash
move to the next field, then the next, then stop.
Fields the real clouds return and these did not, found by comparing each
pack's answer with a recording rather than by a client breaking: eleven on
every Scaleway server product — including capabilities.placement_groups,
which this pack serves and a client checking the capability first would have
read as unsupported — twelve on Outscale's ReadImages, three on ReadVmTypes, seven on Exoscale's template.
Sixteen Exoscale routes that were already served were answering the wrong
shape: instance-type and template inside an instance are bare references {id}, not expanded catalogue entries, and a unit test was locking that
mistake in because it asserted the schema rather than the wire.
A header nobody vouched for is no longer written down. Redaction matched
eight name substrings, and the three dialects served here passed only because
their bearers are called Authorization and X-Auth-Token — a coincidence,
not a rule. Reproduced: X-Auth-Token redacted and X-Consumer carrying the
same value written in full. Request and response headers are now an allowlist;
bodies stay a denylist, because a body is the measurement and a header is
not.
Changed
The Scaleway root volume type has two reasons, and one was missing. The
comment beside rootVolume justified forcing b_ssd with an argument that
covers local volumes only, so it said nothing about sbs_volume — and a
reader lifted the restriction on that basis. docs/limits.md now states the
consequence plainly: no root_volume type is writable today, b_ssd will not
plan from provider 2.79 on, sbs_volume plans for ever, and omitting the
block is the way through. It ends with #8. Reported by Vincent Dislaire.
docs/fourth-pack.md measures what a fourth provider pack would touch:
about 45 additive lines across 13 shared files, and no code in internal/core
naming a provider. The neutral-core rule holds under measurement rather than
by assertion.