It would take the same amount of time to finish with or without turns (concurrency) and you would have done the same amount of work.

But in this case, if you could bring the 8 ex-cashier/cooks/now-cleaners, and each one of them (plus you) could take a zone of the house to clean it, you could do all the work in **parallel**, with the extra help, and finish much sooner.

    tensorflow::ImmediateExecutionTensorHandle* unwrapped_handle =
        tensorflow::unwrap(handles[i]);
    if (tensorflow::CustomDeviceTensorHandle::classof(unwrapped_handle)) {
      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(

  //      since, after all, this class is also a reentrant mutual exclusion lock!?

  /*
   * One of the key challenges of this class is to prevent lost signals, while trying hard to
   * minimize unnecessary signals. One simple and correct algorithm is to signal some other waiter
   * with a satisfied guard (if one exists) whenever any thread occupying the monitor exits the

      "certChain": []
    },
    {
      "identity": "spiffe://cluster.local/ns/istio-system/sa/istiod",
      "state": "Unavailable: signing gRPC error (The service is currently unavailable): error trying to connect: TLS handshake failed: cert verification failed - unable to get local issuer certificate [CERTIFICATE_VERIFY_FAILED]",
      "certChain": []
    }
  ]

So, if your application consumes a lot of memory (for example with machine learning models), and your server has a lot of CPU cores **but little memory**, then your container could end up trying to use more memory than what is available, and degrading performance a lot (or even crashing). 🚨

### Create a `Dockerfile`

Here's how you would create a `Dockerfile` based on this image:

```Dockerfile

tableSize: description: The table size for Maglev hashing. type: integer type: object minimumRingSize: description: Deprecated. type: integer ringHash: description: The ring/modulo hash load balancer implements consistent hashing to backend hosts. properties: minimumRingSize: description: The minimum number of virtual nodes to use for the hash ring. type: integer type: object useSourceIp: description: Hash based on the source IP address. type: boolean type: object localityLbSetting: properties: distribute:...

Pydantic's `orm_mode` will tell the Pydantic *model* to read the data even if it is not a `dict`, but an ORM model (or any other arbitrary object with attributes).

This way, instead of only trying to get the `id` value from a `dict`, as in:

```Python
id = data["id"]
```

it will also try to get it from an attribute, as in:

```Python
id = data.id
```

    ```

!!! info
    FastAPI added support for `Annotated` (and started recommending it) in version 0.95.0.

    If you have an older version, you would get errors when trying to use `Annotated`.

    Make sure you [Upgrade the FastAPI version](../deployment/versions.md#upgrading-the-fastapi-versions){.internal-link target=_blank} to at least 0.95.1 before using `Annotated`.

			attempts := 1
			for attempts <= 3 {
				aerr := api.AbortMultipartUpload(ctx, tgtBucket, tgtObject, res.UploadID, ObjectOptions{})
				if aerr == nil {
					return
				}
				batchLogIf(ctx,
					fmt.Errorf("trying %s: Unable to cleanup failed multipart replication %s on remote %s/%s: %w - this may consume space on remote cluster",
						humanize.Ordinal(attempts), res.UploadID, tgtBucket, tgtObject, aerr))
				attempts++

		// This may be because of 2 reasons:
		//
		// 1) The folder/object was deleted.
		// 2) We come from another disk and this disk missed the write.
		//
		// We therefore perform a heal check.
		// If that doesn't bring it back we remove the folder and assume it was deleted.
		// This means that the next run will not look for it.
		// How to resolve results.
		resolver := metadataResolutionParams{
			dirQuorum: f.disksQuorum,