Worker Protocol:
Note that the worker protocol already had a serialization for
`BuildResult`; this was added in
a4604f1928. It didn't have any versioning
support because at that time reusable seralizers were not away for the protocol
version. It could thus only be used for new messages also introduced in
that commit.
Now that we do support versioning in reusable serializers, we can expand
it to support all known versions and use it in many more places.
The exist test data becomes the version 1.29 tests: note that those
files' contents are unchanged. 1.28 and 1.27 tests are added to cover
the older code-paths.
The keyered build result test only has 1.29 because the keying was also
added in a4604f19284254ac98f19a13ff7c2216de7fe176; the older
serializations are always used unkeyed.
Serve Protocol:
Conversely, no attempt was made to factor out such a serializer for the
serve protocol, so our work there in this commit for that protocol
proceeds from scratch.
This will allow us to factor out logic, which is currently scattered
inline, into several reusable instances
The tests are also updated to support versioning. Currently all Worker
and Serve protocol tests are using the minimum version, since no
version-specific serialisers have been created yet. But in subsequent
commits when that changes, we will test individual versions to ensure
complete coverage.
To start, it is just a clone of the common protocol. But now that we
have the separate protocol implementations, we can add versioning
information without the versions of one protocol leaking into another.
Using the infrastructure from the previous commit, we don't have to
duplicate code for shared behavior.
Motivation: No more perverse incentives. [0] did some awkward things
because the serialisers did not store the version. I don't want anyone
making changes to be pushed towards keeping the serialization logic with
the core data types just because it's easier or the alternative is
tedious.
The actual versioning of the Worker and Serve protocol serialisers
(Common remains unversioned as the underlying mini-protocols are not
versioned) will happen in subsequent commits / PRs.
[0]: fe1f34fa60
This introduces some shared infrastructure for our notion of protocols.
We can then define multiple protocols in terms of that notion.
We an also express how particular protocols depend on each other.
For example, we can define a common protocol and a worker protocol,
where the second depends on the first in terms of the data types it can
read and write.
The "serve" protocol can just use the common one for now, but will
eventually need its own machinary just like the worker protocol for
version-aware serialisers
We want to be able to write down `foo.drv^bar.drv^baz`:
`foo.drv^bar.drv` is the dynamic derivation (since it is itself a
derivation output, `bar.drv` from `foo.drv`).
To that end, we create `Single{Derivation,BuiltPath}` types, that are
very similar except instead of having multiple outputs (in a set or
map), they have a single one. This is for everything to the left of the
rightmost `^`.
`NixStringContextElem` has an analogous change, and now can reuse
`SingleDerivedPath` at the top level. In fact, if we ever get rid of
`DrvDeep`, `NixStringContextElem` could be replaced with
`SingleDerivedPath` entirely!
Important note: some JSON formats have changed.
We already can *produce* dynamic derivations, but we can't refer to them
directly. Today, we can merely express building or example at the top
imperatively over time by building `foo.drv^bar.drv`, and then with a
second nix invocation doing `<result-from-first>^baz`, but this is not
declarative. The ethos of Nix of being able to write down the full plan
everything you want to do, and then execute than plan with a single
command, and for that we need the new inductive form of these types.
Co-authored-by: Robert Hensing <roberth@users.noreply.github.com>
Co-authored-by: Valentin Gagarin <valentin.gagarin@tweag.io>
Pass this around instead of `Source &` and `Sink &` directly. This will
give us something to put the protocol version on once the time comes.
To do this ergonomically, we need to expose `RemoteStore::Connection`,
so do that too. Give it some more API docs while we are at it.
The motivation is exactly the same as for the last commit. In addition,
this anticipates us formally defining separate serialisers for the serve
protocol.
See API docs on that struct for why. The pasing as as template argument
doesn't yet happen in that commit, but will instead happen in later
commit.
Also make `WorkerOp` (now `Op`) and enum struct. This led us to catch
that two operations were not handled!
Co-authored-by: Robert Hensing <roberth@users.noreply.github.com>
This is generally a fine practice: Putting implementations in headers
makes them harder to read and slows compilation. Unfortunately it is
necessary for templates, but we can ameliorate that by putting them in a
separate header. Only files which need to instantiate those templates
will need to include the header with the implementation; the rest can
just include the declaration.
This is now documenting in the contributing guide.
Also, it just happens that these polymorphic serializers are the
protocol agnostic ones. (Worker and serve protocol have the same logic
for these container types.) This means by doing this general template
cleanup, we are also getting a head start on better indicating which
code is protocol-specific and which code is shared between protocols.
This is the more typically way to do [Argument-dependent
lookup](https://en.cppreference.com/w/cpp/language/adl)-leveraging
generic serializers in C++. It makes the relationship between the `read`
and `write` methods more clear and rigorous, and also looks more
familiar to users coming from other languages that do not have C++'s
libertine ad-hoc overloading.
I am returning to this because during the review in
https://github.com/NixOS/nix/pull/6223, it came up as something that
would make the code easier to read --- easier today hopefully already,
but definitely easier if we were have multiple codified protocols with
code sharing between them as that PR seeks to accomplish.
If I recall correctly, the main criticism of this the first time around
(in 2020) was that having to specify the type when writing, e.g.
`WorkerProto<MyType>::write`, was too verbose and cumbersome. This is
now addressed with the `workerProtoWrite` wrapper function.
This method is also the way `nlohmann::json`, which we have used for a
number of years now, does its serializers, for what its worth.
This reverts commit 45a0ed82f0. That
commit in turn reverted 9ab07e99f5.
In many cases we are dealing with a collection of realisations, they are
all outputs of the same derivation. In that case, we don't need
"derivation hashes modulos" to be part of our map key, because the
output names alone will be unique. Those hashes are still part of the
realisation proper, so we aren't loosing any information, we're just
"normalizing our schema" by narrowing the "primary key".
Besides making our data model a bit "tighter" this allows us to avoid a
double `for` loop in `DerivationGoal::waiteeDone`. The inner `for` loop
was previously just to select the output we cared about without knowing
its hash. Now we can just select the output by name directly.
Note that neither protocol is changed as part of this: we are still
transferring `DrvOutputs` over the wire for `BuildResult`s. I would only
consider revising this once #6223 is merged, and we can mention protocol
versions inside factored-out serialization logic. Until then it is
better not change anything because it would come a the cost of code
reuse.
In https://github.com/NixOS/nix/pull/6311#discussion_r834863823, I
realized since derivation goals' wanted outputs can "grow" due to
overlapping dependencies (See `DerivationGoal::addWantedOutputs`, called
by `Worker::makeDerivationGoalCommon`), the previous bug fix had an
unfortunate side effect of causing more pointless rebuilds.
In paticular, we have this situation:
1. Goal made from `DerivedPath::Built { foo, {a} }`.
2. Goal gives on on substituting, starts building.
3. Goal made from `DerivedPath::Built { foo, {b} }`, in fact is just
modified original goal.
4. Though the goal had gotten as far as building, so all outputs were
going to be produced, `addWantedOutputs` no longer knows that and so
the goal is flagged to be restarted.
This might sound far-fetched with input-addressed drvs, where we usually
basically have all our goals "planned out" before we start doing
anything, but with CA derivation goals and especially RFC 92, where *drv
resolution* means goals are created after some building is completed, it
is more likely to happen.
So the first thing to do was restore the clearing of `wantedOutputs` we
used to do, and then filter the outputs in `buildPathsWithResults` to
only get the ones we care about.
But fix also has its own side effect in that the `DerivedPath` in the
`BuildResult` in `DerivationGoal` cannot be trusted; it is merely the
*first* `DerivedPath` for which this goal was originally created.
To remedy this, I made `BuildResult` be like it was before, and instead
made `KeyedBuildResult` be a subclass wit the path. Only
`buildPathsWithResults` returns `KeyedBuildResult`s, everything else
just becomes like it was before, where the "key" is unambiguous from
context.
I think separating the "primary key" field(s) from the other fields is
good practical in general anyways. (I would like to do the same thing
for `ValidPathInfo`.) Among other things, it allows constructions like
`std::map<Key, ThingWithKey>` where doesn't contain duplicate keys and
just precludes the possibility of those duplicate keys being out of
sync.
We might leverage the above someday to overload `buildPathsWithResults`
to take a *set* of return a *map* per the above.
-----
Unfortunately, we need to avoid C++20 strictness on designated
initializers.
(BTW
https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p2287r1.html
this offers some new syntax for this use-case. Hopefully this will be
adopted and we can eventually use it.)
No having that yet, maybe it would be better to not make
`KeyedBuildResult` a subclass to just avoid this.
Co-authored-by: Robert Hensing <roberth@users.noreply.github.com>
More progress on issue #5729
The method trivially generalizes to be store-implementation-agnostic, in
fact.
However, we force it to continue to be unimplemented with `RemoteStore`
and `LegacySSHStore` because the implementation we'd get via the
generalization is probably not the one users expect. This keeps our
hands untied to do it right going forward.
For more about the tension between the scheduler logic being
store-type-agnostic and remote stores doing their own scheduling, see
issues #5025 and #5056.
More progress on issue #5729.
Instead of having it by the default method in `Store` itself, have it be
the implementation in `DummyStore` and `LegacySSHStore`. Then just the
implementations which fail to provide the method pay the "penalty" of
dealing with the icky `unimplemented` function for non-compliance.
Combined with my other recent PRs, this finally makes `Store` have no
`unsupported` calls!
This function returns true or false depending on whether the Nix client
is trusted or not. Mostly relevant when speaking to a remote store with
a daemon.
We include this information in `nix ping store` and `nix doctor`
Co-Authored-By: John Ericson <John.Ericson@Obsidian.Systems>
The references set seems to have been unused since `LegacySSHStore`
references were first created in
caa5793b4a.
The method decls never were upstream, and accidentally added by me in
062533f7cd (probably due to `git rerere`).
Sorry!
This reduces the diff from #3746.
These only functioned if a very narrow combination of conditions held:
- The result path does not yet exist (--check did not result in
repeated builds), AND
- The result path is not available from any configured substituters, AND
- No remote builders that can build the path are available.
If any of these do not hold, a derivation would be built 0 or 1 times
regardless of the repeat option. Thus, remove it to avoid confusion.
I had started the trend of doing `std::visit` by value (because a type
error once mislead me into thinking that was the only form that
existed). While the optomizer in principle should be able to deal with
extra coppying or extra indirection once the lambdas inlined, sticking
with by reference is the conventional default. I hope this might even
improve performance.
This is important if the remote side *does* execute
nix-store/nix-daemon successfully, but stdout is polluted
(e.g. because the remote user's bashrc script prints something to
stdout). In that case we have to shutdown the write side to force the
remote nix process to exit.
Instead of
error: serialised integer 7161674624452356180 is too large for type 'j'
we now get
error: 'nix-store --serve' protocol mismatch from 'sshtest@localhost', got 'This account is currently not available.'
Fixes https://github.com/NixOS/nixpkgs/issues/37287.
When doing e.g.
nix-build -A package --keep-failed --option \
builders \
'ssh://mfhydra?remote-store=/home/bosch/store x86_64-linux - 10 4 big-parallel'
this doesn't work properly because this build-setting is ignored.
I changed this behavior by passing the `settings.keepFailed` through the
serve-protocol to remote machines to make sure that I can introspect the
build-directory (which is particularly helpful when I have to look at a
`config.log` from a failed build for instance).
With this, we don't have to copy the entire .drv closure to the
destination store ahead of time (or at all). Instead, buildPaths()
reads .drv files from the eval store and copies inputSrcs to the
destination store if it needs to build a derivation.
Issue #5025.
This avoids an ambiguity where the `StorePathWithOutputs { drvPath, {}
}` could mean "build `brvPath`" or "substitute `drvPath`" depending on
context.
It also brings the internals closer in line to the new CLI, by
generalizing the `Buildable` type is used there and makes that
distinction already.
In doing so, relegate `StorePathWithOutputs` to being a type just for
backwards compatibility (CLI and RPC).
This is already used by Hydra, and is very useful when materializing
a remote builder list from service discovery. This allows the service
discovery tool to only sync one file instead of two.
