This guide helps data leaders plan and run a Hadoop to Snowflake migration with less risk and fewer surprises.
Hadoop to Snowflake migration is the move of data and workloads from Hadoop clusters to Snowflake’s managed cloud platform, usually in phases, with validation before each cluster retires.
Key Takeaways
- Hadoop to Snowflake migration swaps upkeep-heavy, on-premises Hadoop clusters for a cloud data platform that scales storage and compute separately, and brings analytics and AI under one governance model
- A sound plan starts with discovery, since the inventory of your Hadoop estate decides how long everything else takes
- Phased migrations move reporting workloads first, while Hadoop keeps running until results match on both platforms
- SnowConvert AI converts Hive and Spark SQL code, and Snowpark Connect for Spark runs supported Spark 3.5 DataFrame and SQL jobs
- Validation decides success, so plan parallel runs that compare query results and get business sign-off before every cutover
What Is Hadoop to Snowflake Migration and Why Do Enterprises Do It?
Hadoop to Snowflake migration moves your data and code from self-managed Hadoop clusters to Snowflake, so your team stops running infrastructure and puts its time into analytics.
A Hadoop estate is rarely just HDFS. Most enterprises run a distribution such as Cloudera or MapR, which HPE acquired in 2019, with Hive and Impala for SQL, Spark for processing, Kafka for streaming, Oozie for scheduling and Ranger or Sentry for security. Each piece needs a destination on Snowflake or a clear decision to retire it.
Snowflake changes the operating model. It runs on AWS, Azure or Google Cloud as a managed service, separates storage from compute and bills compute in credits only while virtual warehouses run.
Enterprises usually start the move for one or more of these reasons:
- Hardware refresh: Your cluster is due for new servers, and you’d rather not fund another cycle of on-premises capacity
- Upgrade pressure: Staying current on Hadoop means planned upgrades, such as the Java 17 requirement that arrived with Hadoop 3.5
- Business user demand: Analysts across your organization want fast, governed SQL access without waiting on batch jobs
- AI plans: Your teams want to apply AI to governed data, which Snowflake supports through Cortex AI Functions
- Skills: Engineers who run Hadoop clusters well are getting harder for you to hire and keep
Hadoop to Snowflake Migration Plan: A Step-by-Step Process
A Hadoop to Snowflake migration runs in phases across twelve steps, taking you from setting goals and taking inventory to retiring your last Hadoop cluster and tuning what runs on Snowflake.
The steps below follow the sequence experienced migration teams use. Steps one through six cover discovery and planning. Implementation spans steps seven through nine, while the final steps cover validation and cutover, followed by optimization after go-live.
Step 1: Set Goals and Success Criteria
Before anyone touches data, agree on why you’re migrating and how you’ll know it worked. Goals shape every later decision, from which workloads move first to how much redesign you fund.
Write down success criteria your leadership team can check:
- Cost: A target for total platform cost compared with today’s Hadoop spend on hardware and staff
- Performance: Query response times for your most important dashboards and reports
- Adoption: The number of business users and teams working on Snowflake by a given date
- Decommissioning: The date by which each Hadoop cluster should be switched off
These criteria become the yardstick for every wave and every sign-off that follows.
Step 2: Inventory Your Hadoop Technology Estate
A complete inventory of tools is the foundation of the whole plan. Include tools native to your Hadoop distribution and any third-party software that reads from or writes to the cluster, then reverse engineer a current-state architecture from that list.
| Inventory Category | What to Capture |
|---|---|
| Ingestion and data integration | Tools such as Sqoop, NiFi or Flume, plus any vendor ETL feeding the cluster |
| Processing and transformation | Hive, Spark and MapReduce jobs, split into programmatic jobs and user-facing tools |
| SQL interfaces | Hive and Impala access, from both applications and analysts |
| Data catalog and lineage | Hive Metastore and any catalog or lineage tools in use |
| Visualization | BI tools and dashboards that query Hadoop directly |
| Storage | File formats such as Parquet or ORC, and how much data is structured versus unstructured |
| Streaming | Kafka or MapR Event Store topics, message formats, volumes and integrations |
| Security and governance | Kerberos authentication, Ranger or Sentry policies, HDFS ACLs and encryption settings |
| Third-party software | Any licensed product with a dependency on the cluster |
For each tool, decide if it’s still needed at all. Some get replaced by built-in Snowflake capabilities, while others move to the cloud alongside Snowflake and need their own evaluation for a cloud-native setup.
Step 3: Profile Data Sources and Owners
Data sources deliver the value, so profile them as carefully as the tools. Assign every source an owner or subject matter expert who can answer questions, and use discovery to flag datasets that are no longer used and don’t need to move.
The questions differ by source type:
- Every source: Ownership and classification, plus retention rules and any masking or encryption requirements
- Relational databases: Database type and table counts, along with custom functions and how changes can be captured incrementally
- Streams: Message format and schema registry use, along with topic counts and throughput
- Files: Location and format details, including file encoding and how new data arrives
- APIs: Access method and response schemas, along with request volumes and authentication
How each source handles deletes and schema changes deserves special attention. Those details decide your incremental load design in step seven.
Step 4: Assess Applications and Workload Fit
This is the most important part of discovery. Each application on Hadoop combines several tools, so document how it’s built before you estimate effort.
For every application, document:
- What the application does and the value it provides to the business
- Hadoop technologies it uses, including versions such as Spark 2 or Spark 3
- Batch or streaming mode, plus any external libraries it depends on
- Inputs and outputs, along with every downstream consumer
- Classification of the data it reads and writes
- Support and monitoring setup, including known errors and how they’re resolved
Then sort each workload by fit:
| Workload Type | Fit for Snowflake | What to Plan |
|---|---|---|
| Analytical SQL on tabular data | Strong | Convert Hive or Impala SQL and move it in early waves |
| Spark SQL and DataFrame jobs | Strong to moderate | Check Snowpark Connect for Spark support, which covers Spark 3.5 |
| Spark jobs using RDDs, MLlib or streaming | Needs work | Rewrite in Snowpark or SQL, or keep those jobs on Spark for now |
| Low-latency key lookups on HBase | Evaluate | Test hybrid tables or the search optimization service against your latency targets |
| Unstructured files such as images or video | Partial | Keep files in cloud object storage and access them from Snowflake through stages and directory tables |
Analytical workloads on relational or tabular data are usually the best first candidates.
Step 5: Design the Target Snowflake Architecture
Now design where everything lands. Start with the cloud, since Snowflake runs on AWS, Azure or Google Cloud and the choice shapes storage and identity integration.
Then work through the core design decisions with your architects:
| Decision | What to Settle |
|---|---|
| Account and region | Where your Snowflake account runs, based on data residency and user location |
| Database layout | How databases and schemas map to domains, environments and teams |
| Warehouse design | How many virtual warehouses you need and how to size each one by workload |
| Table strategy | Which data lands in native Snowflake tables and which stays as Apache Iceberg tables in your storage |
| Ingestion pattern | Batch loads with COPY INTO, Snowpipe for files, or Snowpipe Streaming and the Snowflake Connector for Kafka for streams |
| Identity and network | Single sign-on with your identity provider, plus AWS PrivateLink or Azure Private Link on Business Critical edition or higher where policy requires private connectivity |
| Environments | Separate development, test and production setups, with zero-copy clones giving testers realistic data without duplicating storage |
Keep truly unstructured data, such as images and video, in cloud object storage. Snowflake can catalog and process those files through stages and directory tables without loading them into tables.
Step 6: Plan Migration Waves
Group workloads into waves so each one delivers value and lowers risk for the next. A typical sequence looks like this:
| Wave | What Moves | Why |
|---|---|---|
| Foundation | Accounts, roles, networking and the first data loads | Every later wave depends on it |
| Wave 1 | High-value reporting and dashboards on analytical data | Business users see benefits early |
| Wave 2 | Core batch pipelines and their upstream sources | Removes the largest share of cluster load |
| Wave 3 | Complex Spark jobs, streaming and specialist workloads | These need the most design and testing |
| Final wave | Remaining datasets, archives and decommissioning | Lets you switch the cluster off |
Keep dependent workloads in the same wave. Splitting a pipeline from the reports it feeds is one of the easiest ways to stall a cutover.
Step 7: Move Storage and Load Data Into Snowflake
Implementation starts with storage, and this is usually the longest technical step.
- Copy data to cloud storage: Move HDFS data to your cloud provider’s object storage, such as Amazon S3, with Hadoop’s DistCp and the S3A or ABFS connectors
- Keep the folder structure: Mirror the HDFS directory layout one to one, so lineage and existing jobs stay easy to trace
- Re-create storage access: Translate HDFS permissions into your cloud provider’s controls, such as AWS IAM policies or Azure role assignments
- Load into Snowflake: Bulk load historical data with the COPY INTO command, or create Apache Iceberg tables over large datasets that stay in your cloud storage
- Sync ongoing changes: Set up incremental loads with Snowpipe or your existing tools while Hadoop still runs
- Fix small files: Combine small HDFS files before loading, since many tiny files slow loads and queries
Access controls deserve extra time here. Each cloud handles storage permissions differently, so a one-to-one translation of HDFS rules rarely works without redesign.
Step 8: Convert Code and Rebuild Pipelines
With data in place, move the logic. Start with SQL, since it usually converts most cleanly.
- HiveQL and Spark SQL: Convert to Snowflake SQL with SnowConvert AI, which lists Hive and Spark SQL as supported sources, then review the items it flags. Check Impala queries against SnowConvert AI’s current source list, and plan manual review where they differ from HiveQL
- Spark jobs: Run supported Spark 3.5 DataFrame and SQL code through Snowpark Connect for Spark, and rewrite RDD or MLlib code in Snowpark
- MapReduce jobs: Rewrite in Snowflake SQL or Snowpark, and retire jobs nobody uses
- Orchestration: Rebuild Oozie workflows as Snowflake tasks or dbt Projects on Snowflake, or keep your existing orchestrator
- Streaming: Move Kafka feeds to the Snowflake Connector for Kafka with Snowpipe Streaming
Run each converted pipeline against the same inputs as the Hadoop version. Differences are much cheaper to fix here than after cutover.
Step 9: Migrate Security Controls and Integrations
Hadoop security usually spans Kerberos authentication, Active Directory groups, Ranger or Sentry policies, HDFS ACLs and encryption settings, so plan it as a design project.
On Snowflake, authentication moves to single sign-on through your identity provider. Access moves to Snowflake roles organized by team and data domain, while masking policies protect sensitive columns and row access policies limit what each user sees. Horizon Catalog then gives your stewards one place to manage it all.
Integrations need the same care. Applications that connect through JDBC or ODBC can switch to Snowflake’s drivers, and BI tools need their connections and semantic layers repointed and retested.
Step 10: Validate Results at Four Levels
Validation is where trust is earned. Run Hadoop and Snowflake side by side, using A/B testing, until every workload passes four checks:
- Data validation: Run the same query on both platforms and confirm an exact match, starting with row counts and moving to aggregates by key dimensions
- Performance validation: Confirm that critical queries meet the response times you set in step one, using Snowflake’s Query Profile to tune slow ones
- Technical validation: Check that CI/CD pipelines and monitoring work, and that every consuming tool points to Snowflake
- Business validation: Ask report owners to confirm their numbers and approve the cutover
Automate the data checks wherever you can. Manual comparison doesn’t scale across hundreds of tables, while scripted checks can rerun after every load.
Keep a rollback path open throughout this step. Until a workload is signed off, Hadoop remains its system of record, so consumers can switch back quickly if a problem appears.
Step 11: Cut Over and Retire Your Hadoop Clusters
Switch consumers to Snowflake one workload at a time, starting with the waves that passed validation. Freeze changes to the affected Hadoop jobs during each cutover window, and hold a short hypercare period afterward with engineers on hand to fix issues quickly.
Train users before each cutover so they’re ready on day one. People worry about losing tools they know, so show them how their daily work runs on Snowflake, from the queries they write to the dashboards they open each morning.
Once a workload runs cleanly for an agreed period, shut down its Hadoop jobs. When the last wave clears, you can retire the cluster entirely or repurpose its hardware.
Step 12: Optimize Cost and Performance After Go-Live
Cutover isn’t the finish line. The first months on Snowflake are when you set the habits that keep cost and performance where your business case expects them.
| Area | What to Do |
|---|---|
| Warehouse sizing | Review warehouse metering history and resize warehouses that run larger than their workloads need |
| Idle compute | Confirm auto-suspend settings, since suspended warehouses use no credits |
| Spend guardrails | Set resource monitors and budgets so overruns trigger alerts before they hit the invoice |
| Cost attribution | Use query tags and separate warehouses so each team sees its own consumption |
| Query performance | Add clustering keys or the search optimization service only on large tables where queries scan too much data |
| Cleanup | Drop staging tables and unused objects left over from migration waves |
Review these monthly for the first quarter after each wave, then fold them into your regular FinOps routine.
Hadoop to Snowflake Migration Approaches: Choosing Your Path
The four common approaches are lift and shift, an Apache Iceberg move, a full re-architecture and a phased hybrid, and your choice depends on how much you want to change along the way.
| Approach | What It Means | Best When | Trade-Off |
|---|---|---|---|
| Lift and shift | Copy tables and repoint pipelines with minimal redesign | You need to exit hardware or a contract quickly | Old design patterns move with the data |
| Open-format move | Land data as Apache Iceberg tables in your cloud storage and query them from Snowflake | You want data in an open format other engines can read | Needs a clear catalog and governance design |
| Re-architect | Redesign models and pipelines for Snowflake-native features | You want the full benefit of the new platform | Takes longer and needs more design effort |
| Phased hybrid | Move workloads in waves while Hadoop keeps running | Your estate is large or some data must stay on premises | You run two platforms for a period |
Many enterprise programs combine these approaches. A common pattern lifts and shifts cold historical data while re-architecting high-value reporting, with everything else on a phased timeline.
The right mix depends on your goals from step one. If cost and speed matter most, lean toward lift and shift for the first waves. If you’re building toward AI and broad self-service analytics, budget for re-architecture where it pays off most.
Hadoop to Snowflake Technology Mapping: Component by Component
Every Hadoop component maps to a Snowflake feature or a cloud service, and mapping them early shows where conversion is simple and where your team will need deeper redesign.
| Hadoop Component | Snowflake Option | Migration Note |
|---|---|---|
| HDFS storage | Snowflake-managed storage, or Apache Iceberg tables in your cloud storage | Data moves through cloud object storage such as Amazon S3 |
| Hive tables and Metastore | Snowflake tables, or Iceberg tables with a catalog integration | Partitioning may change, since Snowflake manages micro-partitions |
| HiveQL and Impala SQL | Snowflake SQL | SnowConvert AI converts Hive and Spark SQL code |
| Spark jobs | Snowpark, or Snowpark Connect for Spark | Supported Spark 3.5 DataFrame and SQL code runs without rewrites |
| MapReduce jobs | Snowflake SQL or Snowpark | Usually rewritten, since there’s no direct equivalent |
| Oozie and other schedulers | Snowflake tasks or dbt Projects on Snowflake, or your existing orchestrator | Recreate dependencies and schedules |
| Sqoop and NiFi ingestion | Openflow connectors, COPY INTO and Snowpipe | Openflow connectors are built on Apache NiFi |
| Kafka streams | Snowflake Connector for Kafka and Snowpipe Streaming | Rows load continuously with low latency |
| HBase | Hybrid tables or the search optimization service, after testing | Very low-latency transactional lookups may need an operational database |
| YARN resource queues | Virtual warehouses sized per workload | Each team or workload gets independent compute |
| Ranger, Sentry and HDFS ACLs | Roles plus masking and row access policies | Security rules need redesign, not a straight copy |
Most of the mapping is straightforward for analytical workloads. The rows that need the most attention are HBase, Spark code outside Snowpark Connect’s coverage, and security policies.
Snowflake Tools That Support Hadoop to Snowflake Migration
Snowflake offers tools for each part of the move, from SnowConvert AI for SQL conversion to Snowpark Connect for Spark jobs, and its AIM platform now orchestrates whole migrations.
SnowConvert AI
SnowConvert AI is Snowflake’s code conversion tool, and Snowflake documents Hive and Spark SQL as supported source languages, including tables and views. Automated conversion handles most of the syntax, which leaves your engineers to review the items it flags.
Run conversion wave by wave. Each batch then gets reviewed and tested against the Hadoop output while the context is fresh, and lessons from early waves improve the later ones.
Snowpark Connect for Spark
Snowpark Connect for Spark runs supported Spark 3.5 DataFrame and Spark SQL code on Snowflake virtual warehouses. Some Spark capabilities don’t carry over:
| Spark Capability | Status on Snowpark Connect for Spark |
|---|---|
| RDD API and SparkContext | Not available, so rewrite as DataFrame operations |
| MLlib | Not supported, with Snowpark ML or third-party libraries as alternatives |
| Structured Streaming | Not supported, so plan batch or Snowpipe Streaming patterns |
| Spark 4.0 and later | Not supported, since Snowpark Connect targets Spark 3.5 |
Ingestion Tools
Match each Hadoop ingestion flow to the Snowflake option that fits its pattern:
- Openflow connectors: Built on Apache NiFi, a natural landing spot for teams that already run NiFi flows
- Snowpipe: Loads files as soon as they land in a stage, which suits file drops that used to land in HDFS
- Kafka connector: Loads Kafka topics through Snowpipe Streaming with low latency
- Spark connector: Lets existing Spark jobs write straight into Snowflake tables during the transition
Snowflake AIM
Snowflake introduced Snowflake AIM in June 2026 as an AI-powered platform for migration and modernization (Source). Its migration agent tracks project state and progress across the migration lifecycle, and a dedicated AIM desktop application entered public preview in October 2026, walking teams through six stages that start with connecting to the source system. Check its current source support against your Hive and Spark estate before you plan around it.
Hadoop to Snowflake Migration Challenges and How to Solve Them
The hardest parts of a Hadoop to Snowflake migration are hidden job dependencies and code that won’t convert automatically, and both have known remedies, as does cost control after go-live.
| Challenge | How to Solve It |
|---|---|
| Hidden job dependencies | Build lineage during discovery and move dependent workloads in the same wave |
| HiveQL edge cases and custom UDFs | Convert with SnowConvert AI, then assign engineers to the items it flags |
| Spark code built on RDDs or MLlib | Rewrite as DataFrames or Snowpark, or keep those jobs on Spark for now |
| Petabyte-scale data movement | Move cold data in bulk early, then switch to incremental loads |
| Storage access controls | Redesign HDFS permissions for your cloud provider’s identity and access model |
| Small files from HDFS | Combine small files before loading to improve load and query efficiency |
| Uncontrolled credit spend | Set resource monitors and auto-suspend defaults on every warehouse |
| Data that must stay on premises | Keep that data on Hadoop, or in S3-compatible on-premises storage that Snowflake can reach |
Most of these problems show up early if discovery is thorough. That’s why steps two through four deserve real time, even when there’s pressure to start moving data.
Hadoop to Snowflake Migration Cost and Timeline Factors
Cost and timeline depend mainly on data volume and code complexity, along with how many teams consume the data, so scope your estate before you commit to dates or budgets.
Timelines vary widely between enterprises, so we’ve avoided quoting a generic range. These factors drive both cost and duration:
- Data volume and change rate: Larger, faster-changing datasets need longer bulk loads and more incremental sync work
- Code complexity: Custom MapReduce jobs and UDFs take longer than standard HiveQL
- Number of consumers: Every downstream report and application needs repointing and sign-off
- Approach chosen: Lift and shift moves faster, while re-architecture takes longer and returns more value
- Parallel running: Running both platforms adds cost, so set clear exit criteria for each workload
On the Snowflake side, compute bills in credits per second with a 60-second minimum while warehouses run, suspended warehouses use no credits, and storage bills separately. Build that consumption model into your business case alongside today’s Hadoop costs, including hardware refresh and the staff time spent on upgrades.
Hadoop to Snowflake Migration Considerations by Industry
Industry rules shape migration order and validation: regulated sectors move restricted data later and validate longer, while retail and technology teams often move reporting first.
The plan above applies everywhere, yet the sequence and the sign-off process change with your industry.
Financial Services
Banks and insurers usually carry the heaviest regulatory load, so migration timing has to respect reporting calendars. Avoid cutting over regulatory reports near quarter-end or filing deadlines, and bring risk and compliance teams into discovery from the start.
Snowflake roles and access policies map well to strict entitlement models, while Secure Data Sharing gives auditors and subsidiaries live access without file transfers. Data that residency rules keep on premises can stay on Hadoop until policy allows the move. Explore our financial services expertise.
Healthcare and Life Sciences
Protected health data drives every decision here. One practical sequence:
- Move de-identified and aggregated datasets first, so analytics teams gain value early
- Apply masking policies to sensitive columns before identifiable data arrives
- Migrate identifiable records only after privacy and security teams approve the controls
- Keep validation evidence for audits, since reviewers may ask how results were checked
Explore our healthcare and life sciences work.
Retail and CPG
Retailers often hold seasonal change freezes before peak trading periods, so plan cutovers well outside them. Merchant and category dashboards usually make the best first wave, because those users feel the most pain from shared Hadoop clusters.
Consumer goods companies add another layer. Retailer feeds and syndicated data often arrive as files, which suit Snowpipe, and Secure Data Sharing can give retail partners governed access to shared metrics.
Manufacturing
Plants often run batch processing of machine logs close to operations. Keep that processing stable until enterprise KPIs run cleanly on Snowflake, then move plant data in stages, one site or product line at a time. Streaming sensor feeds can shift to the Snowflake Connector for Kafka with Snowpipe Streaming once the core reporting layer is proven.
Technology
Software companies tend to move fastest, since product telemetry and usage analytics map cleanly to Snowflake. Move these workloads early, and use Secure Data Sharing when customers need access to their own usage data without custom exports.
Best Practices for a Successful Hadoop to Snowflake Migration
The best migrations invest in discovery before moving data, keep risk low with waves and parallel runs, and treat cost governance as a day-one habit on Snowflake.
Before You Migrate
- Retire before you move: Drop tables and jobs nobody uses, since every unused object adds conversion and validation effort
- Name an owner for every dataset: Owners answer questions in discovery and sign off in validation
- Agree on success criteria early: Targets for cost and adoption give every later decision a yardstick
- Design governance fresh: Build Snowflake roles and policies around how people use data today, without carrying years of Ranger exceptions forward
While You Migrate
Waves keep risk contained. Pair a high-value reporting workload with a low-risk one in your first wave, so your team builds confidence while business users see benefits early.
Automation matters most for repetitive work. Use SnowConvert AI for Hive and Spark SQL, keep human review for the items it flags, and script your data validation so checks rerun after every load. Zero-copy clones give testers production-like data without paying for duplicate storage.
After Go-Live
| Practice | Why It Matters |
|---|---|
| Resource monitors and budgets on every warehouse | Alerts arrive before overruns reach the invoice |
| Separate warehouses for loading and reporting | Heavy loads don’t slow dashboards for business users |
| Monthly reviews of metering history | Oversized warehouses get caught early |
| Training refreshers for new users | Adoption keeps growing after the migration team moves on |
If you’re also moving other warehouses to Snowflake, our Snowflake migration strategy guide covers the architecture decisions that apply across source systems.
How LatentView Helps With Hadoop to Snowflake Migration
We’ve moved large on-premises Hadoop estates to the cloud. For one client, we planned and executed a Hadoop migration to the AWS cloud with 10x lower TCO, as this case study describes.
Transform your Hadoop estate into a governed Snowflake platform with an end-to-end analytics partner that brings 20+ years of enterprise analytics experience and 50+ Fortune 500 clients. Our MigrateMate solution automates the path from assessment to validation and typically reduces migration costs by 30% to 40%. Start with one workload: talk to our data migration team about a Hadoop assessment.
Frequently Asked Questions
1. How Long Does a Hadoop to Snowflake Migration Take?
Duration depends on data volume and code complexity, along with the number of downstream consumers. A phased program moves the first workload wave sooner, while full decommissioning follows once every workload passes validation.
2. What Are the Main Phases of a Hadoop to Snowflake Migration?
Migration starts with discovery, which inventories the Hadoop estate and sets goals. Implementation then moves data and converts code, and validation compares results on both platforms before each cutover.
3. Can Spark Jobs Run on Snowflake Without Rewrites?
Supported Spark 3.5 DataFrame and Spark SQL code can run on Snowflake through Snowpark Connect for Spark. Jobs built on RDDs or MLlib need a rewrite or should stay on Spark, as do streaming jobs.
4. What Happens to HDFS Data During the Migration?
HDFS data moves to cloud object storage first, usually with Hadoop’s DistCp tool and the same folder structure. From there, Snowflake loads the files with COPY INTO or Snowpipe, or reads the data as Apache Iceberg tables.
5. Can Hadoop and Snowflake Run Together During Migration?
Yes. A phased migration runs both platforms in parallel, comparing query results until each workload matches, and then retires Hadoop jobs one wave at a time.