## Changes
The issue reported in #1828 illustrates how using a YAML timestamp-like
value (a date in this case) causes an issue during conversion to and
from the typed configuration tree.
We use the `AsAny()` function on the `dyn.Value` when normalizing for
the `any` type. We only use the `any` type for variable values, because
they can assume every type. The `AsAny()` function returns a `time.Time`
for the time value during conversion **to** the typed configuration
tree. Upon conversion **from** the typed configuration tree back into
the dynamic configuration tree, we cannot distinguish a `time.Time`
struct from any other struct.
To address this, we use the underlying string value of the time value
when we normalize for the `any` type.
Fixes#1828.
## Tests
Existing unit tests pass
## Changes
Added JSON input validation for CLI commands. Now when invalid JSON
passed as a payload to CLI commands, CLI performs input normalisation
and detects if there are any mismatches such as incorrect types, unknown
fields and etc.
This diagnostic information is printed in standard error output and does
not block command execution, so the change is backward compatible.
Fixes#1769#1764#1625#1560
## Tests
Added unit tests
```
andrew.nester@HFW9Y94129 ~ % databricks jobs create --json '{"seeti}'
Error: error decoding JSON at (inline):1:2: unexpected EOF
andrew.nester@HFW9Y94129 ~ % databricks jobs create --json '{"seeti": true}'
Warning: unknown field: seeti
in (inline):1:9
Error: Job settings must be specified.
```
---------
Co-authored-by: Pieter Noordhuis <pieter.noordhuis@databricks.com>
## Changes
This PR makes sweeping changes to the way we generate and test the
bundle JSON schema. The main benefits are:
1. More modular JSON schema. Every definition in the schema now is one
level deep and points to references instead of inlining the entire
schema for a field. This unblocks PyDABs from taking a dependency on the
JSON schema.
2. Generate the JSON schema during CLI code generation. Directly stream
it instead of computing it at runtime whenever a user calls `databricks
bundle schema`. This is nice because we no longer need to embed a
partial OpenAPI spec in the CLI. Down the line, we can add a `Schema()`
method to every struct in the Databricks Go SDK and remove the
dependency on the OpenAPI spec altogether. It'll become more important
once we decouple Go SDK structs and methods from the underlying APIs.
3. Add enum values for Go SDK fields in the JSON schema. Better
autocompletion and validation for these fields. As a follow-up, we can
add enum values for non-Go SDK enums as well (created internal ticket to
track).
4. Use "packageName.structName" as a key to read JSON schemas from the
OpenAPI spec for Go SDK structs. Before, we would use an unrolled
presentation of the JSON schema (stored in `bundle_descriptions.json`),
which was complex to parse and include in the final JSON schema output.
This also means loading values from the OpenAPI spec for `target` schema
works automatically and no longer needs custom code.
5. Support recursive types (eg: `for_each_task`). With us now using
$refs everywhere it's trivial to support.
6. Using complex variables would be invalid according to the schema
generated before this PR. Now that bug is fixed. In the future adding
more custom rules will be easier as well due to the single level nature
of the JSON schema.
Since this is a complete change of approach in how we generate the JSON
schema, there are a few (very minor) regressions worth calling out.
1. We'll lose a few custom descriptions for non Go SDK structs that were
a part of `bundle_descriptions.json`. Support for those can be added in
the future as a followup.
2. Since now the final JSON schema is a static artefact, we lose some
lead time for the signal that JSON schema integration tests are failing.
It's okay though since we have a lot of coverage via the existing unit
tests.
## Tests
Unit tests. End to end tests are being added in this PR:
https://github.com/databricks/cli/pull/1726
Previous unit tests were all deleted because they were bloated. Effort
was made to make the new unit tests provide (almost) equivalent
coverage.
## Changes
Some call sites hold on to the `dyn.Path` provided to them by the
callback. It must therefore never be mutated after the callback returns,
or these mutations leak out into unknown scope.
This change means it is no longer possible for this failure mode to
happen.
## Tests
Unit test.
## Changes
If not explicitly quoted, the YAML loader interprets a value like
`2024-08-29` as a timestamp. Such a value is usually intended to be a
string instead. Our normalization logic was not able to turn a time
value back into the original string.
This change boxes the time value to include its original string
representation. Normalization of one of these values into a string can
now use the original input value.
## Tests
Unit tests in `libs/dyn/convert`.
## Changes
Since locations are already tracked in the dynamic value tree, we no
longer need to track it at the resource/artifact level. This PR:
1. Removes use of `paths.Paths`. Uses dyn.Location instead.
2. Refactors the validation of resources not being empty valued to be
generic across all resource types.
## Tests
Existing unit tests.
## Changes
This PR:
1. Uses dynamic walking (via the `dyn.MapByPattern` func) to validate no
two resources have the same resource key. The allows us to remove this
validation at merge time.
2. Modifies `dyn.Mapping` to always return a sorted slice of pairs. This
makes traversal functions like `dyn.Walk` or `dyn.MapByPattern`
deterministic.
## Tests
Unit tests. Also manually.
## Changes
Some diagnostics can have multiple paths associated with them. For
instance, ensuring that unique resource keys are used across all
resources. This PR extends `diag.Diagnostic` to accept multiple paths.
This PR is symmetrical to
https://github.com/databricks/cli/pull/1610/files
## Tests
Unit tests
## Changes
This PR changes `diag.Diagnostics` to allow including multiple locations
associated with the diagnostic message. The diagnostics that now return
multiple locations with this PR are:
1. Warning for unknown keys in config.
2. Use of experimental.run_as
3. Accidental sync.exludes that exclude all files.
## Tests
Existing unit tests pass. New unit test case to assert on error message
when multiple locations are included.
Example output:
```
➜ bundle-playground-2 ~/cli2/cli/cli bundle validate
Warning: You are using the legacy mode of run_as. The support for this mode is experimental and might be removed in a future release of the CLI. In order to run the DLT pipelines in your DAB as the run_as user this mode changes the owners of the pipelines to the run_as identity, which requires the user deploying the bundle to be a workspace admin, and also a Metastore admin if the pipeline target is in UC.
at experimental.use_legacy_run_as
in resources.yml:10:22
databricks.yml:13:22
Name: fix run_if
Target: default
Workspace:
User: shreyas.goenka@databricks.com
Path: /Users/shreyas.goenka@databricks.com/.bundle/fix run_if/default
Found 1 warning
```
## Changes
This PR changes the location metadata associated with a `dyn.Value` to a
slice of locations. This will allow us to keep track of location
metadata across merges and overrides.
The convention is to treat the first location in the slice as the
primary location. Also, the semantics are the same as before if there's
only one location associated with a value, that is:
1. For complex values (maps, sequences) the location of the v1 is
primary in Merge(v1, v2)
2. For primitive values the location of v2 is primary in Merge(v1, v2)
## Tests
Modifying existing merge unit tests. Other existing unit tests and
integration tests pass.
---------
Co-authored-by: Pieter Noordhuis <pieter.noordhuis@databricks.com>
## Changes
PyDABs output can omit empty sequences/mappings because we don't track
them as optional. There is no semantic difference between empty and
missing, which makes omitting correct. CLI detects that we falsely
modify input resources by deleting all empty collections.
To handle that, we extend `dyn.Override` to allow visitors to ignore
certain deletes. If we see that an empty sequence or mapping is deleted,
we revert such delete.
## Tests
Unit tests
---------
Co-authored-by: Pieter Noordhuis <pcnoordhuis@gmail.com>
## Changes
The map function ignores cases where either a key in a map is not
present or an index in a sequence is out of bounds. As of recently, we
retain nil values as valid values in a configuration tree. As such, it
makes sense to also ignore cases where a map or sequence is expected but
nil is found. This is semantically no different from an empty map where
a key is not found.
Without this fix, all calls to `dyn.Map` would need to be updated with
nil-checks at every path component.
Related PRs:
* #1507
* #1511
## Tests
Unit tests pass.
## Changes
This PR makes two changes:
1. In https://github.com/databricks/cli/pull/1510 we'll be adding
multiple associated location metadata with a dyn.Value. The Go compiler
does not allow comparing structs if they contain slice values
(presumably due to multiple possible definitions for equality). In
anticipation for adding a `[]dyn.Location` type field to `dyn.Value`
this PR removes all direct comparisons of `dyn.Value` and instead relies
on the kind.
2. Retain location metadata for values in convert.FromTyped. The change
diff is exactly the same as https://github.com/databricks/cli/pull/1523.
It's been combined with this PR because they both depend on each other
to prevent test failures (forming a test failure deadlock).
Go patch used:
```
@@
var x expression
@@
-x == dyn.InvalidValue
+x.Kind() == dyn.KindInvalid
@@
var x expression
@@
-x != dyn.InvalidValue
+x.Kind() != dyn.KindInvalid
@@
var x expression
@@
-x == dyn.NilValue
+x.Kind() == dyn.KindNil
@@
var x expression
@@
-x != dyn.NilValue
+x.Kind() != dyn.KindNil
```
## Tests
Unit tests and integration tests pass.
## Changes
Added support for complex variables
Now it's possible to add and use complex variables as shown below
```
bundle:
name: complex-variables
resources:
jobs:
my_job:
job_clusters:
- job_cluster_key: key
new_cluster: ${var.cluster}
tasks:
- task_key: test
job_cluster_key: key
variables:
cluster:
description: "A cluster definition"
type: complex
default:
spark_version: "13.2.x-scala2.11"
node_type_id: "Standard_DS3_v2"
num_workers: 2
spark_conf:
spark.speculation: true
spark.databricks.delta.retentionDurationCheck.enabled: false
```
Fixes#1298
- [x] Support for complex variables
- [x] Allow variable overrides (with shortcut) in targets
- [x] Don't allow to provide complex variables via flag or env variable
- [x] Fail validation if complex value is used but not `type: complex`
provided
- [x] Support using variables inside complex variables
## Tests
Added unit tests
---------
Co-authored-by: shreyas-goenka <88374338+shreyas-goenka@users.noreply.github.com>
## Changes
This reverts commit dac5f09556 (#1523).
Retaining the location for nil values means equality checks no longer
pass.
We need #1520 to be merged first.
## Tests
Integration test `TestAccPythonWheelTaskDeployAndRunWithWrapper`.
## Changes
There are four different treatments location metadata can receive in the
`convert.FromTyped` method.
1. Location metadata is **retained** for maps, structs and slices if the
value is **not nil**
2. Location metadata is **lost** for maps, structs and slices if the
value is **is nil**
3. Location metadata is **retained** if a scalar type (eg. bool, string
etc) does not change.
4. Location metadata is **lost** if the value for a scalar type changes.
This PR ensures that location metadata is not lost in any case; that is,
it's always preserved.
For (2), this serves as a bug fix so that location information is not
lost on conversion to and from typed for nil values of complex types
(struct, slices, and maps).
For (4) this is a change in semantics. For primitive values modified in
a `typed` mutator, any references to `.Location()` for computed
primitive fields will now return associated YAML location metadata (if
any) instead of an empty location.
While arguable, these semantics are OK since:
1. Situations like these will be rare.
2. Knowing the YAML location (if any) is better than not knowing the
location at all. These locations are typically visible to the user in
errors and warnings.
## Tests
Unit tests
## Changes
With https://github.com/databricks/cli/pull/1507 and
https://github.com/databricks/cli/pull/1511 we are clarifying the
semantics associated with `dyn.InvalidValue` and `dyn.NilValue`. An
invalid value is the default zero value and is used to signals the
complete absence of the value.
A nil value, on the other hand, is a valid value for a piece of
configuration and signals explicitly setting a key to nil in the
configuration tree. In keeping with that theme, this PR returns
`dyn.InvalidValue` instead of `dyn.NilValue` at error sites. This change
is not expected to have a material change in behaviour and is being done
to set the right convention since we have well-defined semantics
associated with both `NilValue` and `InvalidValue`.
## Tests
Unit tests and integration tests pass. Also manually scanned the changes
and the associated call sites to verify the `NilValue` value itself was
not being relied upon.
## Changes
When a configuration defines:
```yaml
run_as:
```
It first showed up as `run_as -> nil` in the dynamic configuration only
to later be converted to `run_as -> {}` while going through typed
conversion. We were using the presence of a key to initialize an empty
value. This is incorrect and it should have remained a nil value.
This conversion was happening in `convert.FromTyped` where any struct
always returned a map value. Instead, it should only return a map value
in any one of these cases: 1) the struct has elements, 2) the struct was
originally a map in the dynamic configuration, or 3) the struct was
initialized to a non-empty pointer value.
Stacked on top of #1516 and #1518.
## Tests
* Unit tests pass.
* Integration tests pass.
* Manually ran through bundle CRUD with a bundle without resources.
## Changes
This came up in integration testing for #1511. One of the tests
converted a `map[string]any` to a dynamic value and encountered a `nil`
and errored out. We can safely return a nil in this case.
## Tests
Unit test passes.
## Changes
Previously, the functions `Get` and `Index` returned `dyn.NilValue` to
indicate that a map key or sequence index wasn't found. This is a valid
value, so we need to differentiate between actual absence and a real
`dyn.NilValue`. We do this with the zero value of a `dyn.Value` (also
captured in the constant `dyn.InvalidValue`).
## Tests
* Unit tests.
* Renamed `Get` and `Index` to find and update all call sites.
## Changes
This change adds support for Lakehouse monitoring in bundles.
The associated resource type name is "quality monitor".
## Testing
Unit tests.
---------
Co-authored-by: Pieter Noordhuis <pcnoordhuis@gmail.com>
Co-authored-by: Pieter Noordhuis <pieter.noordhuis@databricks.com>
Co-authored-by: Arpit Jasapara <87999496+arpitjasa-db@users.noreply.github.com>
## Changes
This PR also fixes empty values variable overrides using the --var flag.
Now, using `--var="my_variable="` will set the value of `my_variable` to
the empty string instead of ignoring the flag altogether.
## Tests
The change using a unit test. Manually verified the `--var` flag works
now.
## Changes
Add `merge.Override` transform. It allows the override one `dyn.Value`
with another, preserving source locations for parts of the sub-tree
where nothing has changed. This is different from merging, where values
are concatenated.
`OverrideVisitor` is visiting the changes during the override process
and allows to control of what changes are allowed or update the
effective value.
The primary use case is Python code updating bundle configuration.
During override, we update locations only for changed values. This
allows us to keep track of locations where values were initially defined
and used for error reporting. For instance, merging:
```yaml
resources: # location=left.yaml:0
jobs: # location=left.yaml:1
job_0: # location=left.yaml:2
name: "job_0" # location=left.yaml:3
```
with
```yaml
resources: # location=right.yaml:0
jobs: # location=right.yaml:1
job_0: # location=right.yaml:2
name: "job_0" # location=right.yaml:3
description: job 0 # location=right.yaml:4
job_1: # location=right.yaml:5
name: "job_1" # location=right.yaml:5
```
produces
```yaml
resources: # location=left.yaml:0
jobs: # location=left.yaml:1
job_0: # location=left.yaml:2
name: "job_0" # location=left.yaml:3
description: job 0 # location=right.yaml:4
job_1: # location=right.yaml:5
name: "job_1" # location=right.yaml:5
```
## Tests
Unit tests
## Changes
We currently issue a warning if an integer is used where a floating
point number is expected. But if they are convertible, we should convert
and not issue a warning. This change fixes normalization if they are
convertible between each other. We still produce a warning if the type
conversion leads to a loss in precision.
## Tests
Unit tests pass.
## Changes
In 0.217.0 we started to emit warning on unknown fields in YAML
configuration but wrongly considered YAML anchor blocks as unknown
field.
This PR fixes this by skipping normalising of YAML blocks.
## Tests
Added regression tests
## Changes
Errors in normalization mean hard failure as of #1319.
We currently allow malformed configurations and ignore the malformed
fields and should continue to do so.
## Tests
* Tests pass.
* No calls to `diag.Errorf` from `libs/dyn`
## Changes
Variable substitution works as if the variable reference is literally
replaced with its contents.
The following fields should be interpreted in the same way regardless of
where the variable is defined:
```yaml
foo: ${var.some_path}
bar: "./${var.some_path}"
```
Before this change, `foo` would inherit the location information of the
variable definition. After this change, it uses the location information
of the variable reference, making the behavior for `foo` and `bar`
identical.
Fixes#1330.
## Tests
The new test passes only with the fix.
## Changes
This adds context to warnings and errors. For example:
* Summary: `unknown field bar`
* Location: `foo.yml:6:10`
* Path: `.targets.dev.workspace`
## Tests
Unit tests.
## Changes
It's not necessary to error out if a configuration field is present but
not set.
For example, the following would error out, but after this change only
produces a warning:
```yaml
workspace:
# This is a string field, but if not specified, it ends up being a null.
host:
```
## Tests
Updated the unit tests to match the new behavior.
---------
Co-authored-by: shreyas-goenka <88374338+shreyas-goenka@users.noreply.github.com>
## Changes
Before this change maps were stored as a regular Go map with string
keys. This didn't let us capture metadata (location information) for map
keys.
To address this, this change replaces the use of the regular Go map with
a dedicated type for a dynamic map. This type stores the `dyn.Value` for
both the key and the value. It uses a map to still allow O(1) lookups
and redirects those into a slice.
## Tests
* All existing unit tests pass (some with minor modifications due to
interface change).
* Equality assertions with `assert.Equal` no longer worked because the
new `dyn.Mapping` persists the order in which keys are set and is
therefore susceptible to map ordering issues. To fix this, I added a
`dynassert` package that forwards all assertions to `testify/assert` but
intercepts equality for `dyn.Value` arguments.
## Changes
While working on #1273, I found that calls to `Append` on a
`dyn.Pattern` were mutating the original slice. This is expected because
appending to a slice will mutate in place if the capacity of the
original slice is large enough. This change updates the `Append` call on
the `dyn.Path` as well to return a newly allocated slice to avoid
inadvertently mutating the originals.
We have existing call sites in the `dyn` package that mutate a
`dyn.Path` (e.g. walk or visit) and these are modified to continue to do
this with a direct call to `append`. Callbacks that use the `dyn.Path`
argument outside of the callback need to make a copy to ensure it isn't
mutated (this is no different from existing semantics).
The `Join` function wasn't used and is removed as part of this change.
## Tests
Unit tests.
## Changes
This change addresses the path resolution behavior in resource
definitions. Previously, all paths were resolved relative to where the
resource was first defined, which could lead to confusion and errors
when paths were specified in different directories. The new behavior is
to resolve paths relative to where they are defined, making it more
intuitive.
However, to avoid breaking existing configurations, compatibility with
the old behavior is maintained.
## Tests
* Existing unit tests for path translation pass.
* Additional test to cover both the nominal and the fallback behavior.
## Changes
We now keep location metadata associated with every configuration value.
When expanding globs for pipeline libraries, this annotation was erased
because of the conversion to/from the typed structure. This change
modifies the expansion mutator to work with `dyn.Value` and retain the
location of the value that holds the glob pattern.
## Tests
Unit tests pass.
## Changes
The new `dyn.Pattern` type represents a path pattern that can match one
or more paths in a configuration tree. Every `dyn.Path` can be converted
to a `dyn.Pattern` that matches only a single path.
To accommodate this change, the visit function needed to be modified to
take a `dyn.Pattern` suffix. Every component in the pattern implements
an interface to work with the visit function. This function can recurse
on the visit function for one or more elements of the value being
visited. For patterns derived from a `dyn.Path`, it will work as it did
before and select the matching element. For the new pattern components
(e.g. `dyn.AnyKey` or `dyn.AnyIndex`), it recurses on all the elements
in the container.
## Tests
Unit tests. Confirmed full coverage for the new code.
## Changes
The `dyn.Path` argument wasn't tested and could regress. Spotted this
while working on related code. Follow up to #1260.
## Tests
Unit tests.
## Changes
This removes the need for the `allowMissingKeyInMap` option to the
private `visit` function and ensures that the body of the visit function
doesn't add or remove values of the configuration it traverses.
This in turn prepares for visiting a path pattern that yields more than
one callback, which doesn't match well with the now-removed option.
## Tests
Unit tests pass and fully cover the inlined code.
## Changes
This change means the callback supplied to `dyn.Foreach` can introspect
the path of the value it is being called for. It also prepares for
allowing visiting path patterns where the exact path is not known
upfront.
## Tests
Unit tests.
## Changes
This change enables the use of bundle variables for boolean, integer,
and floating point fields.
## Tests
* Unit tests.
* I ran a manual test to confirm parameterizing the number of workers in
a cluster definition works.
## Changes
This builds on #1098 and uses the `dyn.Value` representation of the
bundle configuration to generate the Terraform JSON definition of
resources in the bundle.
The existing code (in `BundleToTerraform`) was not great and in an
effort to slightly improve this, I added a package `tfdyn` that includes
dedicated files for each resource type. Every resource type has its own
conversion type that takes the `dyn.Value` of the bundle-side resource
and converts it into Terraform resources (e.g. a job and optionally its
permissions).
Because we now use a `dyn.Value` as input, we can represent and emit
zero-values that have so far been omitted. For example, setting
`num_workers: 0` in your bundle configuration now propagates all the way
to the Terraform JSON definition.
## Tests
* Unit tests for every converter. I reused the test inputs from
`convert_test.go`.
* Equivalence tests in every existing test case checks that the
resulting JSON is identical.
* I manually compared the TF JSON file generated by the CLI from the
main branch and from this PR on all of our bundles and bundle examples
(internal and external) and found the output doesn't change (with the
exception of the odd zero-value being included by the version in this
PR).
## Changes
This is a fundamental change to how we load and process bundle
configuration. We now depend on the configuration being represented as a
`dyn.Value`. This representation is functionally equivalent to Go's
`any` (it is variadic) and allows us to capture metadata associated with
a value, such as where it was defined (e.g. file, line, and column). It
also allows us to represent Go's zero values properly (e.g. empty
string, integer equal to 0, or boolean false).
Using this representation allows us to let the configuration model
deviate from the typed structure we have been relying on so far
(`config.Root`). We need to deviate from these types when using
variables for fields that are not a string themselves. For example,
using `${var.num_workers}` for an integer `workers` field was impossible
until now (though not implemented in this change).
The loader for a `dyn.Value` includes functionality to capture any and
all type mismatches between the user-defined configuration and the
expected types. These mismatches can be surfaced as validation errors in
future PRs.
Given that many mutators expect the typed struct to be the source of
truth, this change converts between the dynamic representation and the
typed representation on mutator entry and exit. Existing mutators can
continue to modify the typed representation and these modifications are
reflected in the dynamic representation (see `MarkMutatorEntry` and
`MarkMutatorExit` in `bundle/config/root.go`).
Required changes included in this change:
* The existing interpolation package is removed in favor of
`libs/dyn/dynvar`.
* Functionality to merge job clusters, job tasks, and pipeline clusters
are now all broken out into their own mutators.
To be implemented later:
* Allow variable references for non-string types.
* Surface diagnostics about the configuration provided by the user in
the validation output.
* Some mutators use a resource's configuration file path to resolve
related relative paths. These depend on `bundle/config/paths.Path` being
set and populated through `ConfigureConfigFilePath`. Instead, they
should interact with the dynamically typed configuration directly. Doing
this also unlocks being able to differentiate different base paths used
within a job (e.g. a task override with a relative path defined in a
directory other than the base job).
## Tests
* Existing unit tests pass (some have been modified to accommodate)
* Integration tests pass
## Changes
When resolving a value returned by the lookup function, the code would
call into `resolveRef` with the key that `resolveKey` was called with.
In doing so, it would cache the _new_ ref under that key.
We fix this by caching ref resolution only at the top level and relying
on lookup caching to avoid duplicate work.
This came up while testing #1098.
## Tests
Unit test.
## Changes
This is a follow-up to #1211 prompted by the addition of a recursive
type in the Go SDK v0.31.0 (`jobs.ForEachTask`).
When populating missing fields with their zero values we must not
inadvertently recurse into a recursive type.
## Tests
New unit test fails with a stack overflow if the fix if the check is
disabled.
## Changes
This feature supports variable lookups in a `dyn.Value` that are present
in the type but haven't been initialized with a value.
For example: `${bundle.git.origin_url}` is present in the `dyn.Value`
only if it was assigned a value. If it wasn't assigned a value it should
resolve to the empty string. This normalization option, when set,
ensures that all fields that are represented in the specified type are
present in the return value.
This change is in support of #1098.
## Tests
Added unit test.
These fields (key and values) needs to be double quoted in order for
yaml loader to read, parse and unmarshal it into Go struct correctly
because these fields are `map[string]string` type.
## Tests
Added regression unit and E2E tests
Pieter Noordhuis
Andrew Nester
shreyas-goenka
Gleb Kanterov
Aravind Segu