What is the difference between spark inferschema and cloudFiles.inferColumnTypes?

[BF7]

We have been using spark.read with inferSchema = True to validate AutoLoader schema inferencing. But I have a suspicion that they do these differently from each other and may not always yield the identical results.

Has anyone ever answered this question? Does anyone know of documentation that can speak to whether there is a difference between them?

[LRALVA · Accepted solution]

Hi @BF7 

Yes — there is a difference between how spark.read(...).option("inferSchema", "true")
and Auto Loader's schema inference (cloudFiles.schemaHints, cloudFiles.inferColumnTypes, etc.) work.
They are not guaranteed to produce identical results,

Key Differences
✅ 1. Inference Timing
spark.read().option("inferSchema", "true"):
Happens immediately, as Spark reads the files in batch.
Schema is inferred from file sample size or first few rows.

Auto Loader:
Uses a schema inference engine behind the scenes.
Can persist schema at cloudFiles.schemaLocation and evolve it.
Not all files are read at once — schema may evolve over time as new fields arrive.

✅ 2. Sampling Behavior
In spark.read, schema inference is based on sample files or rows.
In Auto Loader, it can be configured to infer from fewer or more files, and it tries to do this efficiently.

✅ 3. Data Types
Sometimes Auto Loader infers:
Different numeric types (LongType vs. DoubleType)
Timestamps vs. strings based on pattern matching
Missing fields (from file 1 but present in file 2)
This makes Auto Loader more flexible but less deterministic than batch inferSchema.

✅ 4. Schema Evolution Support
spark.read = no schema evolution
Auto Loader = supports evolving schemas (if cloudFiles.schemaEvolutionMode is enabled)

Docs / References:

https://docs.databricks.com/aws/en/ingestion/cloud-object-storage/auto-loader/schema

https://spark.apache.org/docs/latest/sql-data-sources-json.html#schema-inference-and-evolution

[BF7]

This is fantastic. Thank you so much. Are you familiar with any specific differences in inferring StringType vs. IntegerType?

[LRALVA]

@BF7 

1. Auto Loader is more conservative
It may default to StringType if the field has:
Inconsistent types across files
Mixed nulls and integers
Unexpected characters
This avoids schema evolution conflicts later in streaming

2. spark.read().option("inferSchema", true) is more aggressive
It can more confidently assign IntegerType or DoubleType in batch mode because it:
Samples more of the data at once
Doesn’t have to worry about downstream schema evolution

Example:
[
{ "id": "123" },
{ "id": 456 },
{ "id": "789" }
]
spark.read(..., inferSchema=True) → likely infers IntegerType (casts strings like "123" if parsable)

Auto Loader → likely infers StringType (preserves original types to avoid runtime failures)