roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepare for Snowflake interview questions on how storage and compute are split and how data is loaded, secured and shared. Covers warehouses, micro-partitions, stages, access control and cost tuning.
104 of 104 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 104 new. 0 cut, outside your plan.
Two symptoms that look alike open many Snowflake conversations: one dashboard query that runs for minutes, and a crowd of quick ones queued behind it. A larger warehouse usually helps the first and extra clusters the second, and the reason is the point: compute here is rented, resized and suspended, not installed. Snowflake surfaces in data and analytics engineering loops wherever the company's data already lives in it, inside the design discussion and around the SQL exercise; platform and administrator interviews give it longer, on cost and access.
Most of the questions ask which layer does the work and who pays for it. Why does a repeated query return at once while its warehouse is suspended? Why does a COPY INTO report nothing loaded while yesterday's files still sit in the stage? Why does updating one row rewrite a whole micro-partition? How can a consumer query a table that was never copied into its account, and whose warehouse runs that query?
Interviewers grade these by the point at which an answer comes to rest. Low in the level range it stops at the mechanism: the result cache, the load metadata, files never modified in place. Higher up it stops at a decision with a price: a dashboard refreshing more often than the auto-suspend interval, so the warehouse seldom gets to suspend; a clustering key whose background reclustering costs more credits than its pruning saves; a partner in another region, which a direct share cannot reach without replication.
The first two modules are built as pairs: most mechanisms in Under the Hood return in Core Platform as something Snowflake bills or configures, and each is taught first so that the feature arrives with its cause. Separating storage from compute is the premise of the product, and workload isolation is the problem a warehouse per team solves; the MPP engine around them explains why a bigger warehouse shortens a large scan yet does little for a query stuck behind one skewed join key. The storage topics then describe the table: zone maps skip whole files by their minimum and maximum values, with no index, and immutable files are why an update writes new files and frequent small loads leave fragments.
Core Platform keeps the hub's own sequence. The three layers come first, then the two halves of the bill: virtual warehouses for compute and micro-partitions for storage. Loading through stages follows both, because a bulk load is a warehouse writing micro-partitions, and its advice on file sizes echoes the small-files topic. Data sharing rests on the same split, with the provider keeping the storage and the consumer usually bringing the compute. Cost and performance tuning closes the module, since reading a Query Profile draws on everything above it.
The last two modules ask who may run any of this, then what the SQL itself costs. In Snowflake privileges are granted to roles and roles to users or other roles, so the privilege model leads, GRANT and REVOKE add ownership and future grants, and column masking waits for both, because a masking policy usually decides from the session's role at query time. Query Tuning then weighs materialized views, which trade background refresh credits for cheaper reads, and window functions, where a ROW_NUMBER() deduplication filtered with QUALIFY can meet one oversized partition and spill.
Where to begin depends on what you will do with the account. Analysts and analytics engineers are asked mostly about queries, so Core Platform and Query Tuning come first, with the first module as the reference for causes. Data engineers need the plan whole, loading and cost above all. An administrator hired to own the account reverses the weight: access and cost tuning first, loading and sharing last. With days left, importance and level can shrink the page to the must know questions at the band your posting hires for. With months, study order set on by topics keeps every Core Platform question behind the mechanism it names, and a trial account turns the Query Profile into something you have read rather than heard about: run one query twice, the second time with its warehouse suspended, and compare the two profiles. Turning loaded tables into tested, documented models is outside this plan. On many Snowflake teams that layer is written in dbt; the [dbt roadmap](/roadmaps/dbt) picks the tables up where this plan leaves them.
Where to begin depends on what you will do with the account. Analysts and analytics engineers are asked mostly about queries, so Core Platform and Query Tuning come first, with the first module as the reference for causes. Data engineers need the plan whole, loading and cost above all. An administrator hired to own the account reverses the weight: access and cost tuning first, loading and sharing last.
With days left, importance and level can shrink the page to the must know questions at the band your posting hires for. With months, study order set on by topics keeps every Core Platform question behind the mechanism it names, and a trial account turns the Query Profile into something you have read rather than heard about: run one query twice, the second time with its warehouse suspended, and compare the two profiles.
Turning loaded tables into tested, documented models is outside this plan. On many Snowflake teams that layer is written in dbt; the dbt roadmap picks the tables up where this plan leaves them.