Skip to main content
Skip to main content
Edit this page

Snowflake and ClickHouse: equivalent concepts

The tables below map each Snowflake concept to its ClickHouse equivalent — what to use instead, and where the model differs. For function-by-function SQL syntax mapping, see the Snowflake → ClickHouse SQL translation reference. For the end-to-end migration walkthrough, see Migrating from Snowflake to ClickHouse.

Resource hierarchy

How the platform organizes accounts, logical containers for data, and where compute is provisioned.

SnowflakeClickHouseNotes
OrganizationOrganizationRoot node of the hierarchy in both.
AccountWarehouse — one or more services sharing storageLike a Snowflake account, a ClickHouse warehouse groups multiple compute units that share storage. Each service has its own compute pool but reads and writes from the warehouse's shared storage. Tier and billing are set at the organization level, not the warehouse.
DatabaseDatabaseLogical container for tables. Snowflake uses a Database → Schema → Table hierarchy; ClickHouse flattens this to Database → Table — see Schemas below.

Schemas

A Snowflake schema serves multiple roles and has no single equivalent in ClickHouse. The table below maps each role to its ClickHouse counterpart.

SnowflakeClickHouseNotes
Namespace partitioning — letting objects with the same name coexist (analytics.users vs marketing.users)One database per Snowflake schema, or fold the schema name into the database (analytics.public.eventsanalytics_public.events)Object references move from three-level (DB.SCHEMA.TABLE) to two-level (DB.TABLE).
Logical grouping by domain or processing stage (analytics.raw, analytics.staging, analytics.marts)Separate databases or a consistent naming convention
Permission boundary (GRANT SELECT ON ALL TABLES IN SCHEMA … TO role)SQL grants at the database, table, or column levelDatabase-wide grants cover the schema-level grant footprint; per-table grants are also available for finer-grained control.
Future-grants anchor (GRANT … ON FUTURE TABLES IN SCHEMA …)Database wildcards (GRANT … ON db.* TO role) apply to current and future tablesFuture grants only apply to a whole database — you can't scope them to a subset of tables within it.
Schema OWNERSHIP and MANAGED ACCESS (by default the role with OWNERSHIP holds grant authority on its objects; WITH MANAGED ACCESS centralizes that authority on the schema owner instead)No object ownership in ClickHouse, so grants are always explicit. Mirrors MANAGED ACCESS in Snowflake, but it's the only mode.
Cloning unit (CREATE SCHEMA … CLONE … for environment branching)No copy-on-write at any granularity — see the Storage and tables section for the zero-copy clone rowEvery copy reads source data fully in ClickHouse.
Time Travel and replication boundary (per-schema retention windows, replication policies)Handled at the table level (TTL) or service level (backups, database replication)No intermediate per-schema boundary.
Tagging and classification scope (tags and policies applied at schema, inherited by contained objects)Apply at the table or column levelNo intermediate namespace inherits down.

Roles and access control

ClickHouse Cloud's access layer splits into console roles at console.clickhouse.cloud (organization-, service-, and SQL-console-scoped) for org admin, billing, and service management; and SQL roles and grants inside each service for database, table, and column access.

SnowflakeClickHouseNotes
Account-level system roles (ACCOUNTADMIN, SYSADMIN, SECURITYADMIN, USERADMIN, PUBLIC)Organization roles (Admin, Billing, Org API reader, Member) and service roles (Service admin, Service reader, Service API admin/reader) in the console; SQL roles inside each serviceThe console layer is split across organization-scoped roles (billing, org admin, user management) and service-scoped roles (service config, scaling, backups). SQL roles cover database, table, and column access within a service.
Custom account rolesCREATE ROLE in SQLSame pattern: create a role, grant privileges to it, grant the role to users.
Database roles (a separate role identity scoped to one database)ClickHouse has only one tier of SQL roles, all service-scoped. There's no equivalent to Snowflake's two-tier system of account roles vs database roles.
Role hierarchy (GRANT ROLE … TO ROLE …)GRANT role1 TO role2
Privilege grants on objects (GRANT … ON … TO ROLE …)GRANT … ON db.table TO role
Object ownership and ownership transferAccess in ClickHouse is controlled entirely through explicit grants. Snowflake patterns that rely on owners delegating access need to be rebuilt as explicit role-based grants.
USE ROLE to switch active role per sessionActive roles are set per session via SET ROLE

Compute and capacity

Warehouse terminology

"Warehouse" means different things in the two systems. A Snowflake warehouse is a compute cluster — what runs queries. A ClickHouse warehouse is a grouping of services that share storage and scale compute independently.

How processing is allocated to a query and sized.

SnowflakeClickHouseNotes
Virtual warehouseService compute pool — queries parallelize across all replicasA ClickHouse service runs across one or more replicas (typically 3 by default on Scale/Enterprise); queries parallelize across them. See the callout below for the sizing-model difference.
Warehouse size (XS through 6X-Large)Vertical autoscaling bounds — replica memory in GiB (CPU scales with memory at a fixed 1:4 ratio in standard profiles)Sizing is configured as min/max memory bounds rather than discrete t-shirt sizes; setting min = max effectively fixes the size.
Multi-cluster warehouseManual horizontal scaling — replica count configured via API or consoleBoth add parallel compute to handle concurrency; ClickHouse scales replica count rather than cluster count. ClickHouse Cloud doesn't have a direct equivalent to Snowflake's auto-scaling policies (Standard/Economy) — horizontal scaling is manual today, with autoscaling on the roadmap.
Auto-suspend / auto-resumeService idling — pause when idle, resume on querySame model: compute stops when there's no work, restarts on the next query.
Resource monitors (credit-quota spend caps)Workloads for runtime scheduling; per-query limits (memory, threads, execution time)Partial overlap. ClickHouse workloads cover runtime resource scheduling (memory, CPU/thread allocation, IO, concurrency, priority) but not spend caps — there's no ClickHouse primitive that suspends a service on hitting a credit threshold.
Query Acceleration ServiceNo direct equivalent — service-level autoscaling covers sustained load, not per-query straggler boostsQAS adds serverless compute to outlier queries. ClickHouse has no per-query compute booster; scale the service if queries are consistently large.
Warehouse size vs replica

A Snowflake warehouse size (XS, S, M, …) is a discrete t-shirt that doubles compute at each step. A ClickHouse replica is sized by autoscaling bounds in CPU and memory. Both are the unit of compute allocated to a query — the sizing model differs.

Billing and pricing model

How each platform meters usage and bills it. See ClickHouse Cloud pricing for current rates.

SnowflakeClickHouseNotes
Compute pricing unit — credits (warehouse-size multiplier × runtime)RAM-minutes — metered per minute in 8 GiB incrementsBoth bill compute by time. ClickHouse compute rates vary by tier, region, and cloud provider.
Storage — compressed bytes, includes Time Travel and Fail-safe overheadCompressed bytes; no Time Travel or Fail-safe overheadStorage rates are the same across ClickHouse tiers and vary by region and cloud provider.
Backups — bundled into Time Travel and Fail-safe retention windowsSeparate line on the bill; one backup retained one day by default, configurable per service via backupsClickHouse backups are explicit and configurable, with their own storage line.
Data transfer — public internet egress and cross-cloud transfer chargedPublic internet egress and cross-region data transfer charged; per-tier allowances applyBoth platforms charge for egress and cross-region transfer. Easy to overlook when modeling total cost.
Separately-metered features — Snowpipe, Search Optimization, Auto-clustering, materialized view refresh, replication, Cortex (all metered as "serverless compute" outside warehouse credits)Mostly bundled into service compute — MV refresh, merges, and indexing run on the service's own replicasClickPipes is the explicit exception and is metered separately.
Editions and commitment — Standard / Enterprise / Business Critical / VPS; capacity contractsBasic / Scale / Enterprise tiers; prepaid ClickHouse Credits (CHC) for committed spendSnowflake editions and ClickHouse tiers gate different security and DR features. Both platforms offer committed-spend discounts.

Storage and tables

How tables are stored: engines, schema, partitioning, snapshots, and access primitives.

In ClickHouse, a table's behavior is set at creation time: the engine (MergeTree family) determines merge and storage semantics, and ORDER BY / PARTITION BY / TTL clauses configure physical layout and retention. Many Snowflake per-feature settings map to a clause in the ClickHouse CREATE TABLE statement.

Physical modeling

The mappings below cover the mechanics. Physical schema design often differs between platforms: Snowflake schemas commonly use normalized dimensional models with joins resolved at query time; ClickHouse schemas commonly use denormalized tables sorted by access pattern via ORDER BY, with dictionaries for lookups and materialized views for pre-aggregation. A direct schema port may not perform the same as a schema designed for the target engine.

SnowflakeClickHouseNotes
Permanent tableMergeTree-family tableEngine choice determines storage and merge behavior — pick by access pattern (MergeTree for append-mostly facts, ReplacingMergeTree for upserts, AggregatingMergeTree for pre-aggregations).
Transient table (no Fail-safe)MergeTree tableClickHouse has no Fail-safe tier, so the permanent/transient distinction doesn't apply.
Temporary table (session-scoped)CREATE TEMPORARY TABLESession-scoped temporary tables exist in both; semantics are similar.
External tables3 / gcs / azureBlobStorage table functions for direct file access; Iceberg engine for open catalogsObject storage and open-table formats are read directly through these functions and engines.
Stage (internal / external / user / table)Object storage referenced directly via s3 / gcs / azureBlobStorage table functions; ClickPipes for managed staging on loadClickHouse has no stage object: there's no managed internal storage layer for files awaiting load, and no PUT / GET equivalents for moving files in and out. Read from the bucket directly, or use ClickPipes to coordinate ingest.
Iceberg table (managed or unmanaged)Iceberg engine (read-only)Reads Iceberg tables stored in S3, Azure, HDFS, or local storage; writes aren't supported. See the engine page for the current list of supported features.
Snowflake Open Catalog (Polaris)Iceberg engine with REST catalog supportBoth expose Iceberg tables through a REST catalog; ClickHouse reads from the catalog, ClickHouse itself isn't a catalog server.
Hybrid table (Unistore)ClickHouse is OLAP-only; OLTP-style point reads and writes aren't a supported workload pattern.
Dynamic tableRefreshable MV (scheduled) or incremental MV (per-insert)Dynamic tables maintain a query result on a target lag; ClickHouse MVs cover both the periodic-refresh and per-insert models — see callout under Query model.
Column data type modes (NOT NULL / nullable)Nullable(T) for optional; omit for requiredIn ClickHouse, columns are non-nullable unless wrapped with Nullable(T). Nullability has a small storage and query cost, so use it only when the column actually needs nulls.
VARIANT, OBJECT, ARRAY (semi-structured)JSON, Tuple, Nested, Map, ArrayClickHouse exposes typed alternatives instead of a single variant column — pick the type that matches the data's shape. The JSON type covers schemaless cases; see the SQL translation reference for the full mapping.
Schema evolution (add / drop / modify columns)ALTER TABLE ... ADD / DROP / MODIFY COLUMNSame DDL surface as Snowflake. Many column changes are metadata-only.
Micro-partitionsData parts — created on insert and merged in the backgroundAn implementation detail of how MergeTree organizes rows on disk. Not directly user-controlled.
Clustering keyORDER BY columns in the table definitionDefined as part of the table; data is physically sorted on disk by the ORDER BY columns. ClickHouse sorts at insert time rather than reorganizing in the background.
Data retention (table / database default)TTL clause on the table, column, or partitionBoth support automatic deletion of data older than a configured window. TTL can be set at table creation or via ALTER TABLE ... MODIFY TTL.
Time TravelPoint-in-time backup restore into a new serviceSee callout below — granularity differs significantly.
Fail-safe (7-day Snowflake-only recovery)Recovery beyond the backup window goes through ClickHouse Cloud support, not a self-service tier.
Zero-copy cloneCREATE TABLE ... AS SELECT copy, or backup restore into a new serviceClickHouse has no copy-on-write primitive — every copy reads the source data fully.
Secure viewView with SQL SECURITY DEFINER (runs with the view-owner's privileges)See CREATE VIEW for the syntax and the INVOKER / DEFINER / NONE modes.
Row access policyRow policy — a WHERE-style expression evaluated per userRow policies apply transparently to every query against the table.
SequenceNo direct equivalent — use generateSnowflakeID, generateUUIDv7, or an external generatorClickHouse has no auto-incrementing sequence object; generated IDs are produced per row at insert time.
Time Travel and backups

Snowflake Time Travel is per-table and queryable inline (SELECT … AT (TIMESTAMP => …)), with retention configurable up to 90 days on Enterprise editions. ClickHouse Cloud backups are per-service: restoring creates a new service, historical state isn't queryable inline, and a single table can't be restored back into the original service. These differences are worth noting for workflows that rely on inline per-table point-in-time queries.

Partitioning

Snowflake has no user-controlled partition primitive — micro-partitions are managed automatically. ClickHouse exposes PARTITION BY as an explicit clause, useful for retention (drop a partition) and pruning. There's nothing on the Snowflake side that maps to this directly; clustering keys are the closest user-controlled layout primitive.

Updates and deletes

ClickHouse is append-optimized. There's no SQL MERGE, and ALTER TABLE … UPDATE / DELETE run as background mutations rather than transactional row writes. Update patterns from Snowflake (MERGE, dbt incremental updates) typically port to engine choice in ClickHouse: ReplacingMergeTree keeps the latest row by sort key, CollapsingMergeTree marks deletes inline, and AggregatingMergeTree maintains aggregated state. Engine choice is set at table creation and is non-trivial to change later.

Query model and performance

How queries run and are accelerated — indexes, materialized views, caches, and streaming inputs.

Query acceleration in ClickHouse comes from three layers: primary-key ordering (a sparse index over the on-disk sort order), secondary indexes on non-key columns, and materialized views — incremental or refreshable. The rows below map Snowflake's acceleration features onto these primitives.

SnowflakeClickHouseNotes
Primary key (advisory)Primary key — drives the on-disk sort order and the sparse primary indexNeither system enforces uniqueness; the ClickHouse optimizer uses the key to prune granules, avoid re-sorts, and short-circuit LIMIT.
Foreign key (advisory)Wide tables or dictionaries for lookupsClickHouse doesn't accept foreign-key declarations even as advisory hints.
Search optimization serviceSecondary indexes — bloom-filter, token-bloom, minmaxSame role: accelerate filters on non-key columns. Snowflake's SOS is automatic and uniform across applied columns; ClickHouse asks you to pick the index type per column and tune its parameters.
Cortex Search / Snowflake Cortex SearchFull-text indexToken index over string columns for in-database search.
VECTOR data type and vector searchArray(Float32) plus a vector ANN indexClickHouse has no dedicated VECTOR type — embeddings are stored as Array(Float32) and accelerated with an ANN index for approximate nearest-neighbor lookups.
Materialized viewIncremental MV — updates on each insert into a base tableThe two systems define materialized views differently. Review Snowflake's source-shape requirements and ClickHouse's incremental-MV behavior before porting an existing MV — they aren't a one-to-one swap. Cost is paid at insert time in ClickHouse.
Dynamic tableRefreshable MV — runs the query on a schedule and maintains its result tableDynamic tables target a lag SLA; refreshable MVs run on a cron-style schedule with the same end-state.
Result cacheQuery cacheBoth transparently reuse results of recently executed queries. Snowflake's result cache is service-wide and persistent; ClickHouse's is per-replica and not transactionally consistent.
Task (scheduled SQL)Refreshable MV for query-driven scheduled work; external orchestrator (dbt, Airflow) for procedural pipelinesRefreshable MVs replace the typical task-into-target-table pattern. Snowflake task DAGs (task graphs) have no direct equivalent — model dependencies in the orchestrator.
Stream (CDC over a table)Materialized view over base-table inserts, or ClickPipes for source-side CDCConceptually related but not equivalent: a Snowflake stream tracks change offsets on a table and is consumed by a task or query. A ClickHouse MV reacts on each insert and writes to a destination table. The end-state pattern (react to changes → write) is similar; the offset/consume model isn't.
EXPLAIN / EXPLAIN_JSONEXPLAIN variants (PLAN, PIPELINE, SYNTAX, ESTIMATE)EXPLAIN ESTIMATE reports rows, parts, and marks the query would read; other variants cover deeper plan inspection.
External functions (HTTPS endpoints via API integrations)No direct equivalent — closest options are executable UDFs (local script invocation) or a database engine attaching a live sourceSnowflake external functions invoke remote HTTPS endpoints from inside SQL. ClickHouse has no managed outbound HTTP call from SQL; the surrogates run locally or attach a database, not call an arbitrary service.
Sessions / session variablesPer-statement execution; multi-step state managed in the client or an orchestratorClickHouse has no per-session variables or shared state.

Secondary indexes

Indexes on non-primary-key columns, used when queries filter by columns outside the sort order:

  • Bloom-filter — equality lookups (=, IN)
  • Token-bloom — substring search on tokenized text
  • Minmax — range pruning by per-part min/max
Materialized view update model

ClickHouse has two MV models: incremental MVs update on every base-table insert (cost proportional to the insert), and refreshable MVs run on a schedule. Snowflake materialized views correspond to the incremental model; Snowflake dynamic tables correspond to the refreshable model. Incremental MVs are typically used for high-throughput aggregations; refreshable MVs for periodic snapshots. Source-shape rules differ between platforms — see Snowflake's MV documentation and ClickHouse's incremental MV guide for the per-platform constraints.

Transformation and modeling

How transformation pipelines port over: dbt adapters and the modeling shifts they expose.

SnowflakeClickHouseNotes
dbt on Snowflake (dbt-snowflake adapter)dbt on ClickHouse via the dbt-clickhouse adapterThe adapter covers the standard dbt materializations — view, table, incremental, materialized_view, ephemeral — plus snapshots, seeds, sources, and tests.
dbt incremental (MERGE-based update strategy)dbt incremental — supports append, delete+insert, insert_overwrite, and microbatch strategies (plus a legacy default)ClickHouse incremental models don't issue SQL MERGE; the adapter rewrites the update pattern around append-optimized engines. See the dbt materialization reference for strategy details.
dbt materialized_view (refresh-based)dbt materialized_view — backed by ClickHouse incremental MVs; experimental in the adapterClickHouse MVs update on insert into the base table, not by re-running the model. Source-shape rules differ between platforms — see the materialized_view materialization page.
dbt Clouddbt-clickhouse isn't available in dbt Cloud at the moment; dbt Core is the supported pathdbt Cloud availability is on the roadmap. See the dbt-clickhouse adapter page for current status.
Other transformation frameworks (Coalesce, SQLMesh, etc.)Use the tool's ClickHouse adapterAdapter coverage and maturity vary; verify supported features against the tool's own documentation.

Security and governance

Access control, encryption, masking, and network boundaries.

Secure views and row access policies are listed under Storage and tables. Roles and grants are covered in Roles and access control.

SnowflakeClickHouseNotes
Column masking policies (including tag-based)Column-level grants on specific columns of a table, or data masking patternsGrants apply at the column level. Snowflake's centralized tag/policy governance has no direct equivalent.
Dynamic data masking (function-based)Views, row policies, or function-based transforms — see data masking patternsNo column-mask primitive yet; patterns are SQL-level.
Network policies (IP allowlist)IP allowlists and PrivateLink (AWS / Azure) for ingress restrictionBoth restrict network ingress; ClickHouse adds PrivateLink for private connectivity.
Tri-Secret Secure (customer-managed keys)CMEK on the serviceBYOK in AWS KMS, with rotation and revocation.
Object tagging (governance metadata)ClickHouse exposes metadata via system.* tables rather than user-defined tags.
Data classification (sensitive-data detection)No direct equivalent — external tools (e.g. DataHub)Not a managed feature.
Encryption functions (ENCRYPT / DECRYPT)Encryption functions (encrypt / decrypt)Covers AES-128/256-CBC/GCM and AEAD modes.
OAuth / SAML SSOSSO (SAML, OIDC)Same role; configured in the cloud console.
Audit logs (ACCOUNT_USAGE.LOGIN_HISTORY, QUERY_HISTORY)Cloud audit log and system.query_logBoth systems log admin and query activity.

Data sharing

Cross-organization data exchange and clean-room patterns.

SnowflakeClickHouseNotes
Secure Data SharingRead access to a shared database, or a dedicated service with consumer-specific row policiesClickHouse has no zero-copy cross-account share; sharing uses standard access primitives.
Snowflake Marketplace / ListingsClickHouse has no in-product data marketplace.
Reader accounts (provider-managed consumer)Dedicated service per consumer, or shared service with row policiesSame pattern: isolate consumer access at the service or row-policy level.
Data Clean RoomsRow policies and secure views — assembled per use caseNo managed clean-room product.

Operations and ecosystem

Day-2 concerns: ingestion, ML/BI integration, observability, metadata, and disaster recovery.

ClickHouse surfaces operational state through system.* tables (queries, sessions, replication, parts, metrics) and the cloud console; managed ingestion is handled by ClickPipes; ML, BI, and notebook workflows are typically handled in external systems that read from ClickHouse.

SnowflakeClickHouseNotes
Snowpipe (continuous ingest from object storage)ClickPipes for S3, GCS, and Azure Blob StorageManaged ingest from object storage.
Snowpipe StreamingClickPipes for Kafka, Kinesis, Pub/SubManaged low-latency streaming ingest.
Openflow connectors / Snowflake ConnectorsClickPipes and the broader integrations libraryOpenflow is Snowflake's connector framework, built on Apache NiFi; ClickPipes is ClickHouse Cloud's managed connector platform. Both cover ingest from streaming systems, OLTP sources, and object storage; see each platform's documentation for the current source list.
Kafka connectorClickPipes for Kafka, or the Kafka table engine for self-managed pipelinesSame role; ClickPipes is the managed option.
Snowflake Connector for Postgres / MySQLClickPipes for Postgres, MySQLManaged CDC from OLTP sources.
Snowpark (Python / Java / Scala DataFrames)External Python with clickhouse-connect or another client libraryNo in-database DataFrame runtime; notebook-side libraries cover the same workflow.
Snowpark ML (in-database training)External training and serving (notebooks, Spark, Vertex AI, feature stores) reading from ClickHouse; see AI/ML in Cloud for managed-side featuresClickHouse has no in-database ML — the typical pattern is to use ClickHouse as the analytical store and run training elsewhere.
Cortex LLM functions (CORTEX.COMPLETE, CORTEX.SUMMARIZE, etc.)No in-SQL equivalent — call LLM providers from the application layer or an orchestrator and write results back to ClickHouseSnowflake exposes hosted LLMs as SQL functions. ClickHouse has no in-query LLM functions; Ask-AI in the docs and console is a docs/console helper, not a SQL surface.
Cortex Analyst (natural-language to SQL over your data)Snowflake offers an NL-to-SQL service grounded on your semantic model. ClickHouse has no in-product equivalent.
Snowsight (web UI: editor, dashboards, monitoring, admin)ClickHouse Cloud console, which includes SQL Console, service management, monitoring, and dashboardsThe ClickHouse Cloud console is the equivalent web surface; SQL Console is one component of it, not the whole UI.
Streamlit in Snowflake / Native Apps / Snowpark Container ServicesNo direct equivalent — host Streamlit, container workloads, and packaged apps externally, then query ClickHouse over its native protocol or HTTPClickHouse has no in-product app-hosting, container, or app-distribution layer.
Notebooks in SnowflakeJupyter with clickhouse-connect or another client libraryNo in-product notebook environment; notebook-side libraries cover the same workflow.
INFORMATION_SCHEMANative system.* tables for ClickHouse-specific detail, or the ANSI information_schema views for tool compatibilityBoth surfaces available.
ACCOUNT_USAGE / READER_ACCOUNT_USAGE viewsNative system.* tablessystem.query_log, system.metric_log, system.processes, and othersSame kind of operational telemetry, exposed through system tables.
Query History (UI and view)system.query_log and system.processes for inspection; KILL QUERY to cancelSame information, exposed through system tables instead of a job view.
Data lineage / Snowflake Horizon Catalogsystem.* tables for metadata; external tools (dbt, DataHub) for lineage and qualityClickHouse exposes metadata via system tables rather than a managed catalog product.
Database replication / Account replication / Failover Groups (Snowgrid)Multi-AZ HA within a region (automatic); cross-region replication via Replicated*MergeTree engines or the Enterprise tier's advanced DR featuresSnowgrid is the underlying Snowflake fabric powering cross-region replication, global data sharing, and Failover Groups. On ClickHouse Cloud, Multi-AZ HA is on by default within a region, and cross-region replication is configured per service. Latency between regions affects write performance.