· Updated · 15 min read
How to get data into Snowflake: ingestion methods, tools and costs
The right loading method meets your freshness needs at a total cost you can defend, including Snowflake compute, tool charges, and maintenance.
The loading-service charge is only one part of keeping source data available when the business needs it. You must standardize how sources load and defend the resulting budget. Choosing by advertised price and connector count hides maintenance, downstream compute, and recovery work. Choose by required freshness and the full operating cost instead.
What is Snowflake data ingestion?
Snowflake data ingestion extracts source data and loads it into Snowflake at the freshness you need, with a defined cost and owner. A loading service can handle only the final step. Another component may still need to extract, prepare, and deliver the source data.
The boundary matters because providers and customers can own different parts of the same operation. NIST guidance includes operating, security, compliance, migration, and exit costs when assessing a cloud service. Your comparison must therefore cover the complete path into Snowflake, not only the component that loads the final files or rows.
How do the methods and costs compare?
Compare freshness, ownership, billing units, and limits before advertised prices. The FinOps guidance separates resource-use metrics from fully loaded costs and calls for documented assumptions. The team’s time to build and maintain the pipelines also matters, but you need your own labor assumptions before converting that work into money.
Choose batch, scheduled, or micro-batch delivery, or streaming, based on required freshness. These are different operating choices, not an upgrade sequence. Snowpipe is typically queryable within minutes. Snowpipe Streaming advertises five-second availability as a best case, not a guarantee.
The first table shows the operating choice. The second separates the charges and work that each choice creates.
| Method | Freshness | Who runs it | Billing unit | Main limit |
|---|---|---|---|---|
| COPY INTO docs | Scheduled batch | Your team and Snowflake | Warehouse runtime | Requires prepared staged files |
| Snowpipe overview | Typically minutes | Snowflake serverless service | Loaded GB, defined by format | Requires prepared staged files |
| Streaming overview | As low as five seconds | Producer and Snowflake | Uncompressed input-value GB | Recovery depends on channel mode |
| Openflow costs | Scheduled or connector CDC | Snowflake or your cloud | Pools or active vCPU-hours | Fit depends on deployment |
| Fivetran Standard PAYG | Scheduled or database CDC | Fivetran SaaS | Monthly active rows (MAR) | Salesforce API quota |
| Matillion pricing | Batch or continuous CDC | Matillion or customer runner | Task credits and contract | Cash rates require an Order Form |
| Airbyte billing | Scheduled or database CDC | Airbyte Cloud | Plan and source credits | Source type changes the unit |
| Estuary Cloud PAYG | Continuous CDC; 30-minute default delivery | Managed data plane | Task-hours and data volume | See Estuary’s pricing page |
| dltHub pricing | Scheduled, event-driven, or PostgreSQL CDC | dltHub managed runtime | Plan plus runtime credits | Engineers review pipeline code |
| Method | Snowflake charges | Tool charges | Formula | Team’s time |
|---|---|---|---|---|
| COPY INTO | Warehouse runtime and storage | None for loading tool | Credit rate × billed runtime | Create, stage, and recover files |
| Snowpipe | 0.0037 credits per loaded GB* | None outside Snowflake | Loaded GB × 0.0037 | Create files and manage delivery |
| Snowpipe Streaming | 0.0037 credits per input-value GB | None outside Snowflake | Input-value GB × 0.0037 | Build and operate the producer |
| Openflow | Pools, ingestion, CDC, telemetry | BYOC infrastructure | Active vCPU-hours × 0.0225 credits | Configure flows and source access |
| Fivetran Standard PAYG | Loading warehouse and storage | MAR consumption | Distinct active rows by connection | Configure sources and recovery |
| Matillion Scale | Loading and transformation warehouses | Task credits plus contract fees | Task-hours × task credit rate | Configure pipelines and source access |
| Airbyte Cloud | Loading warehouse and storage | Source credits plus plan fee | Database GB or SaaS rows | Review schemas and custom connectors |
| Estuary Cloud PAYG | Warehouse or Snowpipe Streaming | Published rates | Tasks + captured and delivered volume | Configure endpoints and schemas |
| dltHub commercial plan | COPY, merge, replace, and storage | USD 12,000 plan plus overage | Overage after 5,000 runtime credits | Review code, tests, and source logic |
Snowpipe costs 0.0037 credits per loaded GB as of 13/10/26. Text uses uncompressed size, while binary formats use observed storage size. Snowpipe Streaming costs 0.0037 credits per uncompressed input-value GB as of 12/10/26. Openflow BYOC costs 0.0225 platform credits per active vCPU-hour as of 5/10/26. Airbyte’s database and SaaS units are dated as of 6/10/26. The dltHub plan price is dated as of 27/10/26.
* Pricing and product information come from each vendor's public website and documentation, linked in the text, as reviewed on 30/10/26. Vendors change prices and plans without notice, so check the linked source or contact the vendor for current terms. Where a vendor does not publish a price, the article says so.
Unlike units are not directly comparable without a workload. Rows, transferred bytes, task runtime, runtime credits, and Snowflake warehouse time can produce different bills for the same source volume.
When should you use COPY INTO, Snowpipe, or Snowpipe Streaming?
Start with Snowflake’s own loading services because each source pipeline must eventually place files or rows into Snowflake. Choose based on what the producer supplies and when the data must be available.
COPY INTO
Scope: COPY INTO docs describe loading supported files from internal or external stages. Your upstream process extracts the data, creates the files, and delivers them to the stage.
Execution: A customer-managed warehouse performs the load. Warehouse billing runs per second, with a 60-second minimum whenever a suspended warehouse starts or resumes.
Failures: The ON_ERROR behavior can stop, skip, or continue a load. Duplicate detection becomes uncertain after 64 days. Forced reloads can duplicate rows, while retained staged files support replay.
Snowpipe
Scope: The Snowpipe service loads files that an upstream process has already placed in a supported stage. It does not extract source records or create those files.
Execution: Snowflake-managed serverless compute loads arriving files. The Snowpipe rate is 0.0037 credits per loaded GB as of 13/10/26. New data is typically queryable within minutes, without a latency guarantee.
Failures: Failed files require the documented reload procedure. Processed-file metadata lasts 14 days, and redelivery after that period can create duplicate rows.
Snowpipe Streaming
Scope: The Streaming overview describes SDK and REST producers. Named Channels use offsets and retained source records for recovery. Elastic Channels retain unacknowledged events and provide unordered, at-least-once delivery. Neither mode reads a database transaction log.
Execution: Snowflake-managed resources use volume billing at 0.0037 credits per uncompressed input-value GB as of 12/10/26. Snowflake describes availability “as low as five seconds,” but does not provide an SLA for that figure.
Failures: For Named Channel recovery, reopen the same channel and resume after its committed offset. Replaying the committed record can duplicate it.
Use COPY INTO when you control batch timing and warehouse use. Use Snowpipe for continuing file arrival. Use Streaming when the producer supplies rows and the freshness requirement justifies building and operating the producer. If those rows represent database changes, you must still decide how to capture them.
When do you need CDC into Snowflake?
Use log-based change data capture (CDC) when you must reproduce committed row changes that periodic extraction cannot preserve. Treat capture, the initial snapshot, downstream delivery, and application in Snowflake as separate work.
Transaction logs might not contain the database’s complete history. The DBLog paper combines chunked table reads with continuing log events and watermarks. This mechanism prevents an older snapshot value from following a newer change. It is DBLog’s design, not a claim about every CDC product.
The DBLog design also separates capture from output delivery, so capture does not establish Snowflake query availability. DBLog requires changed-row events in commit order and enough values for its use cases. Product-specific prerequisites can also include source-log access, retention, privileges, and row identity. Budget the initial snapshot, recurring capture, delivery, and destination application separately.
When should you use Openflow and its connectors?
Use Openflow when a documented connector or processor covers the source and its deployment model fits your security and operating requirements.
Openflow
Scope: Openflow includes a Salesforce Bulk API connector. The PostgreSQL connector is generally available. It takes an initial snapshot and then reads WAL changes. Updates and deletes require usable row identity. Missing coverage can use documented Openflow processors or an HTTP flow.
Execution: Snowflake Deployment runs the data plane in Snowpark Container Services. BYOC runs it in your AWS environment. Snowflake Deployment costs include management and runtime pools. BYOC costs include 0.0225 platform credits per active vCPU-hour as of 5/10/26, plus customer cloud and Snowflake costs.
Failures: Persistent storage retains in-flight data while a runtime is replaced, and checkpointed flows resume from committed positions. PostgreSQL operators retain logical-replication and WAL responsibilities. BYOC also adds infrastructure, security, and upgrade work.
When should you choose managed third-party ELT tools?
Choose a managed ELT product when its source behavior, execution location, recovery process, and billing unit fit your operation. Connector count alone does not settle that choice.
Fivetran Standard PAYG
Scope: Fivetran PostgreSQL first copies selected rows and then reads WAL changes through pgoutput. The Salesforce connector supports incremental extraction. Connector SDK covers sources without a maintained connector.
Execution: Standard PAYG runs in Fivetran SaaS. Salesforce applies a rolling 24-hour API quota. Credentials and replicated data pass through Fivetran before Snowflake. MAR billing counts distinct primary keys by connection, table, destination, and month. Repeated changes to one row count once that month.
Failures: Failed syncs follow a retry schedule through day 14 and then pause. After correction, a connection resumes from its stopping point. Re-syncs can repair integrity problems.
Matillion Scale
Scope: PostgreSQL CDC can take a snapshot and then capture inserts, updates, and deletes. Data Loader also handles batch loads. A Custom Connector supports JSON REST APIs when no prebuilt connector exists.
Execution: Full SaaS hosts the runner and secrets vault. Hybrid SaaS, available to Enterprise customers, keeps execution and secrets in the customer environment. PostgreSQL CDC stages Avro in Snowflake and uses a warehouse to merge changes. Scale pricing uses task credits plus Order Form inputs.
Failures: CDC recovery uses WAL checkpoints after the initial snapshot. An interruption during that snapshot restarts it. Restarting a failed Change Log snapshot can duplicate rows.
Airbyte Cloud
Scope: Airbyte PostgreSQL takes an initial snapshot and then reads WAL changes when source prerequisites hold. Salesforce replication supports incremental loads. Connector development can use Builder, YAML components, or Python for some missing source behavior.
Execution: Airbyte security states that Cloud stores credentials in Google Cloud Secret Manager. Payloads pass through Airbyte infrastructure over encrypted channels and are purged after transfer. Source billing uses four credits per database GB or six credits per million SaaS rows as of 6/10/26.
Failures: Breaking schema changes pause affected connections. A lost PostgreSQL replication slot requires a new slot and full resync. Your team retains source permissions, slot health, schema review, and custom connector maintenance.
Estuary Cloud PAYG
Scope: Estuary PostgreSQL backfills existing tables and then captures inserts, updates, and deletes. Its Salesforce connector captures historical and continuing data. The connector reference documents packaging custom connectors as containers.
Execution: Collections retain captured documents for replay. The Snowflake connector stages changes and applies them transactionally. Its documented default 30-minute cadence supports hourly delivery, but it does not guarantee query availability. Current rates are on Estuary’s pricing page.
Failures: Recovery logs retain checkpoints. Snowflake writes are transactional, and documented retries do not duplicate received rows. Failing shards raise alerts.
When should you run code-first pipelines on dltHub?
dltHub fits a team that wants inspectable Python and managed operations while retaining responsibility for source logic, tests, review, and deployment choices.
dltHub commercial plan
Scope: Start with source requirements and review the generated pipeline before deployment. The ingestion workflow creates inspectable Python, runs a limited sample, and checks rows, schemas, types, pagination, and cursor movement. Engineers enter credentials and validate business correctness. Source interfaces cover custom extraction when no maintained source exists.
The runtime prepares JSONL, Parquet, or CSV files. The Snowflake destination uses internal stages and PUT by default, with one table stage per destination table. S3 and Google Cloud Storage staging are also documented. Snowflake executes COPY, merge, and replace work.
PostgreSQL can use logical replication. The source snapshots existing rows, then handles inserts, updates, and hard deletes when the table has usable row identity.
Execution: dltHub runtime usage and Snowflake charges are separate. The pipeline runtime extracts and normalizes records, creates files, transfers them, and manages state. Snowflake runs COPY, merge, and replace operations. Credentials are encrypted and scoped to a workspace or profile. SaaS jobs run in a regional dltHub VPC, while BYOC jobs run in the customer’s cloud.
Failures: Preserve completed work during recovery. The production behavior retains retry state, resumes pending packages, and commits incremental state after a package loads. Schema changes can retry or follow contracts. PostgreSQL updates and deletes still require usable row identity.
Budget for the commercial plan and both compute boundaries. The plan costs USD 12,000 per month as of 27/10/26. It includes 5,000 credits measured in small-instance runtime hours and has a 12-month minimum. Runtime credit use equals execution hours multiplied by the documented machine factor. Per-job instance sizing is currently in preview, as of 27/10/26. Snowflake credits and storage remain separate.
Choose by freshness, ownership, and cost
Define the required freshness and name the owner for extraction, loading, monitoring, and recovery. Then apply each documented formula to your volumes, rows, task runtime, managed runtime, warehouse runtime, and service plan. Keep the team’s time qualitative unless your company supplies its own hours and labor rates.
The defensible choice meets the freshness requirement, includes each charge it creates, and has a named operator.
Snowflake ingestion FAQ
What’s the cheapest way to load data into Snowflake?
No method has the lowest complete cost without a workload. COPY INTO billing depends on warehouse runtime and leaves file creation to your team. Snowpipe and Snowpipe Streaming each charge 0.0037 credits per applicable GB, but they use different GB definitions and require different upstream work.
Snowpipe or Snowpipe Streaming?
The Snowpipe service loads staged files and typically makes data queryable within minutes. The Streaming service accepts rows from a producer and advertises availability as low as five seconds without an SLA. Choose based on whether your producer supplies files or rows.
Can Snowpipe do CDC?
The Snowpipe service does not read a source transaction log. It can load files containing captured changes, but another component must capture and represent those changes for Snowflake to apply.
How does ingestion differ from transformation, orchestration, and reverse ETL?
Data ingestion moves source data into a staging area or target store. Data transformation changes data structure or content. Workflow orchestration schedules and coordinates tasks. Reverse ETL sends warehouse data to operational systems.
Book a call to try dltHub with a Snowflake ingestion pipeline.
What is dltHub?
The agentic data layer on top of any data warehouse. Agents build, deploy and heal your ingestion on dltHub's managed infrastructure.
Was this page helpful?