README.dev: Document the blessings

Signed-off-by: Ian Jackson <Ian.Jackson@eu.citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
---
v2: Improvements from review

diff --git a/README.dev b/README.dev
index 65ec1112ffc91dccc2441cf009e80ebf17447dcd..351cd25a91a091673719344d19ef7596dc2d07e6 100644 (file)
--- a/README.dev
+++ b/README.dev
@@ -232,3 +232,113 @@ probably are after a reboot) by looking for processes relating to the
 lists flight,job combinations and then:
 
 $ ./mg-hosts previoustasks --clear
+
+
+Flight blessings and the blessed-* host flags
+=============================================
+
+Both flights and hosts have a `blessing'.
+
+Blessings are used:
+
+ * To avoid accidentally tools operating on flights which are in the
+   wrong phase of their existence.
+
+ * To control which automatic archaeologists will look at which
+   flights.  (Archeaology tools take arguments to control which
+   blessings they will consider.)
+
+ * To control which hosts are considered suitable for use with which
+   flight.  (Flights with blessing `foo' use only hosts which have
+   flag `blessed-foo'.)
+
+Each flight has a `blessing' (flights.blessing in the DB) and an
+`intended blessing' (flights.intended).  The `blessing' changes as the
+flight goes through the stages of its lifecycle.  The `intended
+blessing' is roughly what the flight is going to be blessed as when
+its execution has completed, and controls host allocation.
+
+There is also the blessing which the flight will actually get when it
+is finished, which is not represented in the DB; rather, it is the
+blessing argument to sg-execute-flight (aka the -B argument to
+mg-execute-flight).  Normally this is the same as the intended
+blessing in the DB, except for bisection flights.
+
+These are the principal (intended) blessings:
+
+ * `real': Flights which are fully production runs.  They use only
+   clean working trees with no `funny' environment variables, and
+   versions of osstest intended for production.  The data from such
+   flights is used, where applicable, to control production push gate
+   decisions, etc.
+
+   `blessed-real' hosts are fully production-ready.
+
+ * `play': Playground.  Flights may use weird software, dirty working
+   trees, strange environment variables, and so on.  It does not make
+   sense to point archaeology tools at such flights.
+
+   `blessed-play' hosts may be partially or completely nonfunctional,
+   or strange in some way.  For this reason `play' flights should not
+   normally use the automatic host allocator.
+
+ * `adhoc': Special-purpose flights.  They should be run with software
+   nearly as clean as `real': the version of the software used may not
+   be a branch intended for-production, but any enhancements or
+   modifications should not leave false or misleading information in
+   the history database.
+
+   `blessed-adhoc' hosts are just as good as `blessed-real' ones.
+
+   The `adhoc' blessing exists mostly to protect the main production
+   workflow from unintentional violations (for example, unexpectedly
+   buggy historical flight data).
+
+ * `commission-<something>': Flights used for testing that a
+   particular subset of the hosts are working properly.  The hosts to
+   be tested should be blessed `commission-whatever' during
+   commissioning, and that blessing removed and replaced with `real'
+   when the hosts are ready.
+
+ * `real-bisect' and `adhoc-bisect': These are found only as the
+   blessing of finished flights.  (This is achieved by passing
+   *-bisect to sg-execute-flight.)  This allows the archaeologist
+   tools to distinguish full flights from bisection steps.
+
+   The corresponding intended blessing (as found in the `intended'
+   column of the flights table) is `real'.  So the hosts used by the
+   automatic host allocator are those flagged `blessed-real' or
+   `blessed-adhoc' - there are no separate blessed-*-bisect hostflags.
+
+There are also blessings for unfinished flights:
+
+ * `constructing': The flight metadata (jobs, runvars, etc.) is still
+   being populated.  The tools for flight construction (and flight
+   construct editing) generally insist that they operate only on
+   flights in this state.
+
+ * `running': At least one of the jobs in this flight has started.
+   Flight execution software will generally (perhaps via standard
+   library calls) mark a flight `running' if they find it
+   `constructing', and refuse to operate on other flights.
+
+ * `broken': Something went catastrophically wrong.
+
+ * `broken-*': Conventionally set by hand when a flight needs to be
+   `un-blessed', perhaps because it contains data misleading to the
+   archaeologists.
+
+   A flight's blessings can be manually changed with cs-flight-bless:
+      ./cs-flight-bless FLIGHT broken-real real
+   updates FLIGHT to be marked `broken' rather than `real'.
+   (cs-flight-bless matches the existing blessing against its final
+   argument, to avoid accidents.)
+
+Note that osstest is generally `crash-only software': nothing will
+`clean up' a crashed flight and set its blessing to `broken'.  Flights
+which were in progress once upon a time but never completed, because
+they crashed, are simply left with whatever blessing they had at the
+time.
+
+There is a special exception to the tools' flight status checks: any
+flight whose blessing contains `play' can be operated on out of order.