* or after it has returned. If startup fails, the invocation should cause a call to {@link
   * #notifyFailed(Throwable)} instead.
   *
   * <p>This method should return promptly; prefer to do work on a different thread where it is
   * convenient. It is invoked exactly once on service startup, even when {@link #startAsync} is
   * called multiple times.
   */
  @ForOverride
  protected abstract void doStart();

  /**

		Joins("Manager").
		Preload("Manager.Tools").
		Where("users.name Like ?", "TestMergeNestedPreloadWithNestedJoin%").
		Find(&result).Error

	if err != nil {
		t.Fatalf("failed to preload and find users: %v", err)
	}

	AssertEqual(t, result, users)
	AssertEqual(t, len(query), 2) // Check preload queries are merged

	if !regexp.MustCompile(`SELECT \* FROM .*tools.* WHERE .*IN.*`).MatchString(query[0]) {

# JSON Compatible Encoder

There are some cases where you might need to convert a data type (like a Pydantic model) to something compatible with JSON (like a `dict`, `list`, etc).

For example, if you need to store it in a database.

For that, **FastAPI** provides a `jsonable_encoder()` function.

## Using the `jsonable_encoder`

Let's imagine that you have a database `fake_db` that only receives JSON compatible data.

TF_CAPI_EXPORT unsigned char TF_GetXlaConstantFoldingDisabled();
TF_CAPI_EXPORT void TF_SetXlaConstantFoldingDisabled(
    unsigned char should_enable);

// Create a serialized tensorflow.ConfigProto proto, where:
//
// a) ConfigProto.optimizer_options.global_jit_level is set to ON_1 if
// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.

#       options in "$@", and eventually passing that to Java.
#
#       Where the inherited environment variables (DEFAULT_JVM_OPTS, JAVA_OPTS,
#       and GRADLE_OPTS) rely on word-splitting, this is performed explicitly;
#       see the in-line comments for details.
#
#       There are tweaks for specific operating systems such as AIX, CygWin,
#       Darwin, MinGW, and NonStop.
#

   * were obtained by calling {@code snapshot()} on the same {@link PairedStatsAccumulator} without
   * adding any values in between the two calls, or if one is obtained from the other after
   * round-tripping through java serialization. However, floating point rounding errors mean that it
   * may be false for some instances where the statistics are mathematically equal, including

# ==============================================================================
#
# Changes the behavior in pycpp.sh from "run all tests" to "verify that all
# tests can compile." Used in some CI jobs (macOS and Linux Arm64) where test
# execution is too expensive.
TFCI_PYCPP_SWAP_TO_BUILD_ENABLE=1

### Outbound

Requests leaving a pod go through the "outbound" code path on port 15001.
This is where most of Ztunnel's logic lives.

For outbound traffic, we need to first determine where the traffic is destined to.
As Ztunnel operates at L4, we only have the destination IP/port (recovered via `SO_ORIGINAL_DST`).

	ErrUnexpectedEOF = DecoderError{msg: "unexpected end of JSON input"}
	ErrMaxDepth      = DecoderError{msg: "maximum recursion depth exceeded"}
)

type errPos [2]int // line number, byte offset where error occurred

// DecoderError contains a detailed decoding error.
type DecoderError struct {
	msg       string // description of error
	context   string // additional error context
	pos       errPos
	atChar    byte

// {
//   "version": "1",
//   "format": "XXXXX",
//   "XXXXX": {
//
//   }
// }
// Here "XXXXX" depends on the backend, currently we have "fs" and "xl" implementations.
// formatMetaV1 should be inherited by backend format structs. Please look at format-fs.go
// and format-xl.go for details.

// Ideally we will never have a situation where we will have to change the
// fields of this struct and deal with related migration.