Azure Interview Questions for Data Engineers
Azure data engineering interviews dig into storage: which tier to use, how folders really work, and how to grant access safely. This guide prices the tiers with Azure's live rates for Central India and runs the real Azure Storage SDK against Microsoft's Azurite emulator to show prefixes, overwrite protection and SAS.
GloryTecks Team
Training & Placement Faculty
Last updated: 9 October 2026
Azure interview questions for data engineers tend to start with services, such as Data Factory, Databricks, Synapse and Fabric, but the questions that separate candidates are about storage: which tier to use and what it really costs, how folders work, and how to give a partner or a job access without handing over the keys. This guide answers those with numbers and real calls. Prices come live from Azure's public Retail Prices API (Azure: Retail Prices API). Storage operations use the real Azure Storage SDK against Azurite, Microsoft's local storage emulator (Microsoft Learn: Azurite), with one month of New York taxi trips as the data.
import io
from datetime import datetime, timedelta, timezone
import pandas as pd
import requests
from azure.core.exceptions import HttpResponseError, ResourceExistsError
from azure.storage.blob import BlobClient, BlobSasPermissions, BlobServiceClient, generate_blob_sas
print("A1 storage tiers: live retail prices for Central India, locally redundant storage (USD)")
api = ("https://prices.azure.com/api/retail/prices?$filter=serviceName eq 'Storage' and armRegionName eq 'centralindia' "
"and productName eq 'General Block Blob v2' and priceType eq 'Consumption'")
items = requests.get(api, timeout=30).json()["Items"]
price = lambda meter: next(i["retailPrice"] for i in items if i["meterName"] == meter and i["tierMinimumUnits"] == 0)
for tier in ("Hot", "Cool", "Cold", "Archive"):
stored = price(f"{tier} LRS Data Stored")
read_back = 0.0 if tier == "Hot" else price(f"{tier} Data Retrieval" if tier != "Cold" else "Cold LRS Data Retrieval")
print(f" {tier:<8} 1 TB stored for a year {stored * 1024 * 12:7.2f} | reading 1 TB back once {read_back * 1024:6.2f}")
# Azurite, Microsoft's local storage emulator, with its documented development account
AZURITE = ("DefaultEndpointsProtocol=http;AccountName=devstoreaccount1;"
"AccountKey=Eby8vdM02xNOcqFlqUwJPLlmEtlCDXJ1OUzFT50uSRZ6IFsuFq2UVErCz4I6tq/K1SZFPTOtr/KBHBeksoGMGw==;"
"BlobEndpoint=http://127.0.0.1:10000/devstoreaccount1;")
service = BlobServiceClient.from_connection_string(AZURITE)
container = service.get_container_client("trips")
if container.exists():
container.delete_container()
container.create_container()
print("\nA2 'folders' in flat blob storage are name prefixes")
trips = pd.read_parquet("yellow_tripdata_2024-01.parquet", columns=["tpep_pickup_datetime", "total_amount"])
trips = trips[(trips.tpep_pickup_datetime >= "2024-01-01") & (trips.tpep_pickup_datetime < "2024-02-01")]
for day, rows in trips.groupby(trips.tpep_pickup_datetime.dt.date):
buffer = io.BytesIO()
rows.to_parquet(buffer, index=False)
container.upload_blob(f"year=2024/month=01/day={day.day:02d}/trips.parquet", buffer.getvalue())
print(f" blobs under year=2024/month=01/: {len(list(container.list_blobs(name_starts_with='year=2024/month=01/')))}"
f" | under year=2024/month=01/day=1: {len(list(container.list_blobs(name_starts_with='year=2024/month=01/day=1')))}")
print("\nA3 uploading the same name again without overwrite")
try:
container.upload_blob("year=2024/month=01/day=15/trips.parquet", b"duplicate")
except ResourceExistsError as error:
print(" refused:", error.error_code)
print("\nA4 shared access signatures: least privilege and expiry")
name = "year=2024/month=01/day=15/trips.parquet"
key = service.credential.account_key
def sas_client(permission, expiry):
token = generate_blob_sas("devstoreaccount1", "trips", name, account_key=key, permission=permission, expiry=expiry)
return BlobClient(account_url=service.url, container_name="trips", blob_name=name, credential=token)
soon = datetime.now(timezone.utc) + timedelta(hours=1)
reader = sas_client(BlobSasPermissions(read=True), soon)
print(f" read-only token, read: {len(reader.download_blob().readall()):,} bytes")
for label, client in (("read-only token, write", reader),
("expired token, read", sas_client(BlobSasPermissions(read=True), datetime.now(timezone.utc) - timedelta(minutes=5)))):
try:
client.upload_blob(b"x", overwrite=True) if "write" in label else client.download_blob().readall()
except HttpResponseError as error:
print(f" {label}: refused with {error.status_code} {getattr(error.error_code, 'value', error.error_code)}")
container.delete_container()A1 storage tiers: live retail prices for Central India, locally redundant storage (USD)
Hot 1 TB stored for a year 245.76 | reading 1 TB back once 0.00
Cool 1 TB stored for a year 135.17 | reading 1 TB back once 10.24
Cold 1 TB stored for a year 55.30 | reading 1 TB back once 30.72
Archive 1 TB stored for a year 24.58 | reading 1 TB back once 24.58
A2 'folders' in flat blob storage are name prefixes
blobs under year=2024/month=01/: 31 | under year=2024/month=01/day=1: 10
A3 uploading the same name again without overwrite
refused: BlobAlreadyExists
A4 shared access signatures: least privilege and expiry
read-only token, read: 575,417 bytes
read-only token, write: refused with 403 AuthorizationPermissionMismatch
expired token, read: refused with 403 AuthorizationFailureA1. How do you choose a storage tier?
By how often the data is read, how fast it must come back, and how long it stays. At the Central India rates returned on the day of the run, a terabyte kept for a year cost about 246 USD in hot, 135 in cool, 55 in cold and 25 in archive. But cheaper storage means paid reads: reading that terabyte back once cost nothing extra from hot, about 10 USD from cool, 31 from cold and 25 from archive. Archive is also offline, with rehydration taking hours. And each cooler tier has a minimum stay, 30 days for cool, 90 for cold and 180 for archive, with an early deletion charge if you leave sooner (Microsoft Learn: access tiers).
The interview answer: recent, frequently read data hot; finished months cool or cold; data kept only for compliance in archive, governed by a lifecycle policy rather than by hand.
A2. Are folders in blob storage real?
Not in a flat namespace: a "folder" is just a shared prefix in the blob name. The 31 daily files were stored as year=2024/month=01/day=NN/trips.parquet, and listing by prefix found 31 for the month. Listing day=1 returned 10, days 1 and 10 to 19, because a prefix is matched character by character, not as a folder. Zero-padding and a trailing slash avoid the surprise. With a hierarchical namespace, as in Azure Data Lake Storage, directories are real objects, and renaming or deleting one is a single atomic operation instead of one call per blob (Microsoft Learn: hierarchical namespace). For analytics pipelines that rename output folders, that matters.
A3. How do you stop a pipeline from silently overwriting data?
Do not overwrite by default. Uploading to an existing blob name without the overwrite option was refused with BlobAlreadyExists, so a rerun cannot quietly replace yesterday's file. Make overwriting an explicit choice in the code, or write each run to a new path and switch readers over, which also makes reruns easy to reason about.
A4. How do you give limited, temporary access?
With a shared access signature: a token that grants specific permissions on a specific resource until an expiry time (Microsoft Learn: SAS). A read-only token downloaded the 15 January file, 575,417 bytes, but an upload with the same token was refused with 403 AuthorizationPermissionMismatch. A token that had expired five minutes earlier was refused with 403 AuthorizationFailure. Least privilege and short expiry limit the damage if a token leaks. For Azure services talking to each other, managed identities avoid secrets entirely and are usually the better answer.
A5. Which Azure data services should you know?
Data Factory for orchestration and copying, Databricks for Spark and Delta Lake, and Microsoft Fabric, which brings data engineering, warehousing and Data Factory together and is where Microsoft's current data engineering certification, DP-700, is focused (Microsoft Learn: Fabric Data Engineer Associate). Interviewers expect you to say when you would use each, and to know that Fabric is where new Microsoft data work is heading.
A6. Where do pipeline secrets belong?
Not in code, notebooks or pipeline parameters. Store connection strings and keys in Azure Key Vault, and let services such as Data Factory or Databricks reach storage through a managed identity with a role assignment, so no secret exists to leak. When a secret is unavoidable, reference it from Key Vault at run time and rotate it on a schedule. The emulator key used in the code above is a published development key, which is exactly why it must never appear in a real pipeline.
How to answer Azure questions
- Price the whole life of the data, storage and reads, not the storage rate alone.
- Prefer real directories (ADLS) for analytics layouts, and zero-pad partition names.
- Grant the least access for the shortest time, and prefer managed identities.
Pipeline questions are covered in ADF pipelines explained and copying on-premises SQL to Azure with ADF. For the other clouds, see AWS interview questions for data engineers and GCP interview questions for data engineers; for the full round, data engineering interview questions and answers.
What are common Azure interview questions for data engineers?
What is the difference between Blob Storage and ADLS Gen2?
When should I use the archive tier?
What is a SAS token?
Cloud data engineering is part of the Data Engineering course.
Sources
- Azure Retail Prices API (Microsoft) — Microsoft
- Access tiers for blob data (Microsoft Learn) — Microsoft
- Azure Data Lake Storage hierarchical namespace (Microsoft Learn) — Microsoft
- Grant limited access with shared access signatures (Microsoft Learn) — Microsoft
- Use the Azurite emulator for local Azure Storage development (Microsoft Learn) — Microsoft
- Microsoft Certified: Fabric Data Engineer Associate (Microsoft Learn) — Microsoft
GloryTecks Team
Training & Placement Faculty
The GloryTecks faculty team mentors learners across Data Science, AI and Data Engineering, with a focus on real-time projects and placement readiness in Hyderabad.
Want to master Interview Questions?
Learn Data Engineering at GloryTecks in Hyderabad, with trainers who have industry experience, real-time projects and placement assistance.
Related articles
Python Interview Questions for Data Roles
The plain-Python coding tasks data-analyst and data-engineer interviews use, without pandas: counting, deduplicating, grouping, streaming large files, flattening JSON, joining, and the bugs interviewers plant. Each has a worked solution and the follow-up questions to expect.
Data Science Interview Questions and Answers 2026
A data science interview is several interviews in one: statistics, machine learning, coding, product sense and a take-home. This guide maps each round with model answers, then works a churn take-home on real telecom data, including the checks that separate a strong submission from a high score.
AWS Interview Questions for Data Engineers
AWS interview questions for data engineers on S3, Glue, Athena, Redshift, Kinesis, Step Functions and security, each with the points a strong answer includes, the follow-up to expect, and three end-to-end scenario questions.