CCNA Model Deployment Questions

Logging the private wheel alongside the model and referencing it with a relative pip path in the conda environment file lets the serving build resolve the dependency from the model artifacts themselves, avoiding external network access. This is the documented pattern for private libraries and keeps the environment reproducible and self-contained for deployment.

The MLflow client exposes `set_registered_model_alias`, which updates the alias pointer in the registry. When the client is configured with the Databricks tracking URI and the service principal's credentials, the call moves `champion` to the validated version, and the serving endpoint that resolves that alias picks it up. This is the supported programmatic promotion path.

Re-logging with an explicit signature and input_example ensures the logged schema matches the payload the endpoint expects, and registering a new version plus updating the served entity points the endpoint at the corrected model. This directly addresses the signature mismatch reported in the 400 error.

The MLflow Model Registry is the purpose-built service in Databricks for managing the entire model lifecycle. It allows for versioning, metadata tracking, and stage transitions, providing an organized approach to moving models from initial experimentation through staging to final deployment in a production serving endpoint.

Registering the model to Unity Catalog under a three-level namespace like `catalog.schema.model_name` makes it discoverable and governed, and provides the versioned artifact that Model Serving references. This is the prerequisite step before creating an endpoint. Without a registered Unity Catalog model version, the serving configuration has no valid entity to load.

Databricks Model Serving supports traffic splitting on an existing endpoint through the `traffic_config` field, where each entry in `routes` specifies a served model name and a percentage. Setting version 8 to 10% and version 7 to 90% preserves the same endpoint URL and load-balances live requests accordingly, enabling safe canary-style evaluation without touching client code.

The Champion model is intended to be the standard version that applications and users should interact with. It serves as the primary version for production workloads, ensuring that all traffic is routed to the most stable and performant model version currently vetted by the organization's governance processes.

Scale-to-zero lets the endpoint release all compute when no requests arrive, which directly satisfies the requirement to scale down to zero replicas when idle. When traffic resumes, the endpoint scales back up, incurring a cold start. Because the scenario demands scaling to zero, this setting is necessary; without it the endpoint would keep at least one replica running and continue consuming capacity during idle periods.

Inference tables persist the request and response payloads of the endpoint to a Delta table, which can be queried continuously. Comparing logged feature values against training statistics in SQL and raising an alert on the drift metric gives an automated trigger for retraining, and it requires no change to the model artifact. This directly matches the requirement for a non-invasive drift safeguard.

Databricks Secrets allow you to reference sensitive information securely during deployment. This approach ensures that credentials remain encrypted at rest and in transit, and are only accessible by authorized users or service principals, aligning with enterprise security standards for cloud-based machine learning deployment workflows and infrastructure.

Inference tables log request and response payloads, and the serving layer extracts predictions based on the expected response structure. If the model returns a dict or array that does not match the expected prediction field, the endpoint can still return HTTP 200 while the logged prediction column is null. This mismatch between the model's output shape and the logging schema is a common cause of missing predictions in inference tables.

Inference tables automatically log request payloads and response payloads from a Model Serving endpoint into a Delta table in Unity Catalog. Enabling them satisfies the requirement to capture every request and response for monitoring and debugging, and the logged data can be queried directly with SQL for drift analysis or troubleshooting.

Model Serving provisions a managed, autoscaling HTTP endpoint for a registered model and returns predictions from JSON request payloads. It requires no infrastructure work, handles scaling and versioning, and supports the request-response pattern the external application needs. Creating the endpoint from the Unity Catalog model is the direct way to obtain a REST API.

Databricks Model Serving directly consumes a registered model version URI, builds the serving environment from the logged conda dependencies, and exposes a REST endpoint with managed scaling. It requires no cluster or server administration, matching the requirement for minimal operational overhead while providing low-latency inference for the registered model.

Inference logging requires a target Delta table where the logs will be written. The table must be specified in the endpoint configuration, and the endpoint's service principal must have write permissions to it. This table stores the request and response data for monitoring.

The error message explicitly points to a schema mismatch where a 'string' was provided instead of a 'double'. This is a direct violation of the model signature, which governs the expected input data types. The client application must be updated to cast the data correctly before sending the request.

Inference tables automatically log request and response payloads from a Model Serving endpoint into a Delta table in Unity Catalog. Each row records the timestamp, request, response, and metadata, enabling downstream joins with ground-truth labels and monitoring of data drift or model quality. Enabling inference tables is the supported way to capture payloads for later analysis.

Inference tables can be configured to log metadata only, and payloads that exceed the maximum logged size are truncated or omitted. When status codes are 200 but payload columns are empty, the most likely cause is that payload logging is disabled in the endpoint configuration or that requests exceeded the size threshold, leaving only metadata rows.

Logging inference inputs and predictions to Delta tables creates an audit trail that enables monitoring. This data can be analyzed to measure drift in input features or model outputs. This is the industry-standard approach in Databricks for building a feedback loop that informs when a model needs to be updated.

Inference tables are a built-in Model Serving feature that automatically logs the request payloads and model responses to a Delta table in Unity Catalog. Enabling them requires no extra pipelines or clusters and directly satisfies the auditing requirement, making it the least-infrastructure solution for capturing endpoint traffic.

Databricks Model Serving supports multiple served entities on one endpoint with configurable traffic percentages. Adding a served entity for version 4 and setting its traffic share to 10% achieves the canary rollout while version 3 continues to receive the remaining traffic.

Staging environments are designed to replicate the production environment as closely as possible. This allows for testing the entire pipeline, including data connectivity, API endpoints, and model performance, ensuring that any issues are detected and resolved before the model is exposed to actual production traffic and users.

Inference tables are stored as Delta tables in Unity Catalog or the Hive metastore, and the endpoint's UI exposes the fully qualified table name. Users can query them with SQL or Spark and join to downstream label tables using the request ID column that Databricks includes. This enables drift analysis and model monitoring. The table contains request payloads, responses, and metadata such as timestamps and model version.

The MLmodel file is generated by MLflow and includes a signature field that defines the input and output schema. Model Serving reads this to validate and parse incoming requests, ensuring the JSON payload matches the expected columns and types. This enables automatic request validation and correct inference.

Model Serving builds the environment from the dependency specification captured with the model. If a package is missing or pinned to a version that cannot be resolved, the build fails. Correcting the `requirements.txt` or aligning `conda.yaml` entries with versions available in the Databricks runtime lets the builder resolve the environment while still installing the libraries the model needs, preserving the model's required dependencies.

HTTP 429 responses from Model Serving typically indicate that the per-replica request queue is saturated even though average CPU is low, because bursts arrive faster than replicas can accept them. Raising `max_instances` allows more replicas to spin up, and tuning the concurrency target makes the autoscaler react sooner, absorbing bursts before requests are rejected.

MLflow records dependencies when the model is logged, and Model Serving rebuilds the environment from those captured requirements. Adding the library to the model's conda environment or pip requirements at log time ensures the endpoint installs it during container build. This is the supported, reproducible mechanism for custom dependencies in served models.

Databricks Model Serving supports Unity Catalog alias-qualified model paths. Pointing the endpoint at `model_v3@champion` causes the serving infrastructure to resolve the alias at request time, so when a new version is promoted to the `champion` alias, the endpoint automatically serves the updated artifact without endpoint reconfiguration or downtime.

Databricks Model Serving endpoints expose a stable HTTPS URL that clients can call with bearer-token authentication. Using a dedicated service principal scoped to query permissions satisfies authentication without granting configuration rights, honoring least privilege. This gives the external application a durable, callable endpoint while ensuring it cannot alter the endpoint or reach unrelated workspace resources, which is exactly what the scenario requires.

Model Serving requires a logged MLflow model artifact it can load, and a model signature defines the expected input and output schema. Without a signature, the endpoint cannot validate incoming request payloads, so a properly logged model with a signature is a genuine prerequisite for successful endpoint creation.

Databricks Model Serving supports `traffic_config` with a `routes` array, where each entry names a served entity and specifies an integer `traffic_percentage`. Summing to 100 across routes enables canary or A/B routing. Setting 95 for champion and 5 for challenger delivers the desired split. This is the supported mechanism for gradual rollout without redeploying the endpoint.

Setting min_instances to 1 keeps one replica provisioned at all times, so the container and model are already loaded and the first request after an idle period is served without a cold start. This is the documented way to trade a small amount of always-on compute for predictable low latency, and it does not require scaling the endpoint beyond what the workload needs during quiet periods.

The signature specifies the required input features and the output format. This metadata is used by the serving endpoint to validate incoming JSON requests, ensuring they match the expected schema. This prevents invalid data from causing failures during the inference process, which is critical for production stability.

The Serving tab in the model version UI provides an interactive interface where a user can send a sample request to the model and view the response. It also displays the model signature, including input and output schema, so the data scientist can confirm expected fields before promoting the version to production.

Model Serving reads the MLflow signature stored with the registered model version and uses it to validate incoming payloads against the expected DataFrame schema. Requests whose columns or types do not match are rejected before inference runs, satisfying the requirement without extra configuration.

Endpoints move through provisioning and updating states, and during those transitions the service may reject requests because no ready replica can serve them. Once the endpoint reaches Ready, traffic is accepted. This is the most direct explanation for an existing endpoint returning a not-ready error, and it clears without any client-side change.

Including a requirements.txt file or letting MLflow capture the environment ensures that the serving environment mirrors the training environment. This is the standard method to maintain consistency, allowing the Databricks serving infrastructure to install the correct package versions during the initial container build and deployment process.

Loading a specific numeric version pins the job to an immutable artifact, guaranteeing that reruns months later use the identical weights and preprocessing. This is the standard way to achieve reproducibility for batch scoring, and it complements the signature metadata. Relying on a moving alias would instead silently change behavior whenever the alias is reassigned, breaking reproducibility.

The model artifact size and its runtime memory consumption directly dictate the minimum compute resources needed. If the model requires more RAM than the chosen workload size provides, the endpoint will fail to load or experience frequent crashes. Assessing memory usage during the testing phase is critical for size selection.

The model signature provides the necessary schema metadata for the serving endpoint. This allows Databricks to enforce input validation for all incoming REST API requests. By defining the signature during the log_model call, you ensure the serving container knows how to translate JSON payloads into the correct format.

Inference tables capture the request payload, the response, and associated metadata into a Delta table in Unity Catalog. Enabling this on the endpoint gives the fraud team an auditable record of each scored transaction that can be joined later with confirmed fraud labels. This is the native, governed mechanism for payload-level observability on Model Serving endpoints.

Model Serving endpoints are protected by Databricks authentication and Unity Catalog or workspace permissions. Granting the caller permission and supplying a bearer token ensures requests are authenticated and authorized. This combination is the supported way to expose an endpoint to an external application while enforcing access control, satisfying both the reachability and the authorization requirements in the scenario.

Provisioned Throughput reserves dedicated resources for the model, which eliminates the variability associated with shared serverless infrastructure. This results in consistent, predictable performance and lower latency, making it the ideal choice for production applications that must handle high-volume traffic with strict performance and reliability requirements.

Model Serving GPU workload sizes require that the workspace has GPU instances enabled and sufficient quota. If GPU instances are not available or the quota is exceeded, the endpoint deployment will fail. This is a common configuration issue when attempting to use GPU-accelerated serving.

This feature facilitates the collection of inference inputs and outputs into a structured format, such as a Delta table. This allows data scientists to monitor the model's performance in production, conduct drift analysis, and use the data for future fine-tuning or retraining efforts, which is vital for maintaining model accuracy.

This is the correct approach because Databricks Lakehouse Monitoring can compute custom metrics from the endpoint's inference table, and alerts can trigger a Databricks job. That job can then update the endpoint configuration to revert to the previous model version. This provides the required automatic, threshold-based rollback within the specified time frame using native Databricks capabilities.

Keeping a minimum instance count above zero holds provisioned, model-loaded replicas ready at all times. Requests arriving during business hours are handled immediately rather than waiting for a container to start and load artifacts, which removes the cold-start delay. This is the standard control for latency-sensitive endpoints that must respond predictably even after periods of low traffic.

Aliases are mutable pointers to a model version, and a served entity can reference an alias instead of a fixed version. Moving the alias from the old version to the new one shifts traffic without editing the endpoint, and moving it back reverts instantly. This matches the requirement to promote and roll back without configuration changes.

MLflow automatically logs the environment dependencies, including Python packages and versions, when log_model is called. The serving infrastructure reads this metadata to rebuild the environment in the container. This ensures that the code runs in an environment identical to the one used during training and testing phases.

Updating the served entity to the new model version URI performs a managed rolling deployment while preserving the ability to revert by re-pointing to the earlier version. This keeps the change reversible and uses the endpoint's native configuration, matching both the traffic redirection and rollback requirements without extra infrastructure.

Passing an `input_example` gives MLflow a concrete sample to infer the input schema and store it alongside the model. Even when an explicit signature is used, an input example helps document expected payloads and can be used to validate the schema. Together with the signature, it ensures the endpoint knows the correct column names and types. This is a recommended practice for reliable deployments.

Keeping min_instances at 1 preserves a warm replica that can answer requests immediately, eliminating scale-from-zero cold starts while max_instances still caps cost during spikes. This balances the cost-saving goal during idle periods with the latency goal during bursts, because only one instance runs continuously rather than four.

Keeping at least one instance warm avoids provisioning delay during bursts, directly reducing cold-start latency. Retrying transient failures with exponential backoff in the calling application or model code handles the intermittent feature store slowness. Together these settings address both the latency and reliability concerns described, making this the balanced operational choice.

When a model is logged without an explicit signature, MLflow attempts to infer it from the input example or from the model's internal schema, which can be incomplete or incorrect. The Serving endpoint relies on that logged signature to validate and parse incoming requests, so the engineer should check whether the signature was explicitly provided and correct. Supplying a proper `ModelSignature` during logging ensures the endpoint enforces the expected column names and types.

Setting min_instances above zero keeps at least that many replicas provisioned and ready, so requests arriving after an idle period do not wait for a new container to load the model. This directly addresses cold-start latency during bursts, at the cost of paying for the always-on capacity. It does not by itself record request payloads or predictions.

The correct administrative path is to update the Access Control List (ACL) for the specific model object. A workspace administrator or the current owner has the authority to grant the required 'CAN_MANAGE' permission to the user, allowing them to perform the requested stage transition securely and officially.

Serving endpoints can target a model alias rather than a fixed version. Pointing the endpoint at an alias like 'champion' lets the team repoint the alias to a new version, and the endpoint picks up the change without altering the URL. This decouples consumers from version numbers and supports controlled promotion.

Endpoint event logs and metrics expose provisioning, build, and update events along with latency and error rates. These reveal whether the failure is a deployment or configuration problem, such as a failed model build or a scaled-to-zero event, complementing request-level payload inspection from inference tables.

Serving endpoints are designed for low-latency, synchronous request-response cycles. They expose the model via an API, allowing applications to request a prediction and receive it in milliseconds. This is the defining requirement for real-time applications where end-user interaction is waiting for the model result.

Logging the model with a signature records the expected input and output schema in the MLflow model artifact. Model Serving reads that signature to validate request payloads, so this is the correct way to enable schema validation and prevent malformed inputs from reaching the model.

Databricks Model Serving creates endpoints from registered model versions, referencing a three-level Unity Catalog name such as catalog.schema.model and a version number. Registering the logged model makes the artifacts, signature, and environment metadata available for the serving layer to build and deploy. Without a registered version, there is no model entity the endpoint can point to, so registration is the required prerequisite.

Unity Catalog models are referenced by their fully qualified three-level name, catalog.schema.model_name, so prod.ml_team.fraud_detector is exactly the identifier Model Serving expects. Because the model lives in Unity Catalog rather than the workspace registry, the endpoint resolves the model version directly from that namespace, and permissions on the catalog, schema, and model govern who can serve it.

Loading the model with the MLflow predict API inside a Spark job lets the scoring run distributed across the cluster, so throughput scales with the data volume and executor count. The job can be scheduled with a Databricks job and retried on failure, and results are written back to Delta for dashboards. This aligns with a nightly, re-runnable bulk scoring pattern.

Databricks Model Serving authenticates as a platform-managed identity and requires explicit Unity Catalog grants on the model's parent catalog, schema, and the model itself. Without USE CATALOG, USE SCHEMA, and EXECUTE on `prod.ml.churn_model`, the endpoint cannot fetch artifacts and creation fails. Granting these to the serving identity or the requesting user resolves the deployment.

Inference tables provide a permanent, queryable record of every inference call, including the input features and the resulting prediction. This is the foundation for monitoring, drift detection, and debugging. By logging this data to Delta tables, you gain full insight into how the model is being used in production.

Databricks Model Serving lets you serve a specific registered model version by supplying the model name and version number in the served entity configuration. By specifying version 3 and avoiding aliases or the `latest` marker, the endpoint stays pinned to that immutable version even as newer versions are registered, which directly satisfies the requirement of always serving the same model.

Serving a model by name plus an alias lets Model Serving resolve the alias at request time, so promoting a new version to `champion` updates the served model without endpoint edits. This matches the requirement of automatic pickup on promotion and works with Unity Catalog-registered models, provided the endpoint's identity has EXECUTE on the model and its versions.

Databricks Model Serving integrates with Unity Catalog so that an endpoint can run under a service principal identity. By granting that principal `EXECUTE` on the model and `SELECT` on feature tables, the endpoint accesses resources without embedding credentials. This satisfies both the no-long-lived-credentials policy and the need for a dedicated identity. The service principal's Unity Catalog privileges govern all access at runtime.

Databricks Model Serving supports configuring multiple served entities on one endpoint and assigning each a percentage of traffic. By giving the new model version a small percentage and the current production model the remainder, the team can perform a canary-style rollout and compare live behavior before a full switch. This is exactly the built-in mechanism for gradual, controlled deployment of a new model version.

Inference tables capture the request and response payloads of a serving endpoint into a Delta table automatically, with no client changes. Teams can then join those logs against training data to compute drift metrics and retain an audit trail. Enabling the feature is a configuration step on the endpoint, matching the requirement for automatic, code-free logging.

Databricks Model Serving provides managed REST endpoints for registered models and supports scale-to-zero, which shuts down compute when there is no traffic and cold-starts on the next request. This satisfies both the native management and idle-cost requirements. The endpoint exposes a standard REST interface and integrates with Unity Catalog for permissions and lineage.

Databricks Model Serving accepts pandas DataFrame inputs in JSON using the dataframe_records or dataframe_split format, where each record maps column names to values. This aligns with the signature logged for a scikit-learn or similar model and lets the endpoint validate and score the record without custom parsing.