* Negative caching: don't perform a build if a dependency already

failed in a previous build.  This is essential for Nixpkgs: we don't
  want to keep doing the same failed dependency (say, Glibc) over and
  over again for a few hundred jobs.

src/sql/hydra.sql

@@ -62,6 +62,7 @@ create table BuildResultInfo (
     --   2 = build of some dependency failed
     --   3 = other failure (see errorMsg)
     --   4 = build cancelled (removed from queue; never built)
+    --   5 = build not done because a dependency failed previously
     buildStatus   integer,
 
     errorMsg      text, -- error message in case of a Nix failure
@@ -74,18 +75,22 @@ create table BuildResultInfo (
     releaseName   text, -- e.g. "patchelf-0.5pre1234"
 
     keep          integer not null default 0, -- true means never garbage-collect the build output
+
+    -- If buildStatus == 5, the primary key of the failed build step.
+    failedDepBuild  integer,
+    failedDepStepNr integer,
     
     foreign key   (id) references Builds(id) on delete cascade -- ignored by sqlite
 );
 
 
 create table BuildSteps (
-    id            integer not null,
+    id            integer not null, -- !!! rename to "build"
     stepnr        integer not null,
 
     type          integer not null, -- 0 = build, 1 = substitution
 
-    drvPath       text, 
+    drvPath       text,
     outPath       text,
 
     logfile       text,