Home › AWS › AWS data services › Direct Connect, VPN, Transit Gateway

AWS data services · Security, identity and network · about 12 min read

AWS Direct Connect, AWS Site-to-Site VPN and AWS Transit Gateway

The private roads from plants, offices and SAP into your AWS VPCs: a leased line, encrypted tunnels over the internet, and one hub joining them.

RoleSecurity, identity and network
StatusGenerally available. Site-to-Site VPN added large bandwidth tunnels (up to 5 Gbps) and a VPN Concentrator for many small sites (announced Nov 2025); Cloud WAN is the managed global alternative to a Transit Gateway.
Verdict on this sitePursue
AzureExpressRoute, VPN Gateway and Virtual WAN
Google CloudCloud Interconnect, Cloud VPN and Network Connectivity Center
Databricksnone: Databricks relies on your cloud network; classic compute sits in your VPC behind it
Open sourcestrongSwan or a router image for IPsec; your own MPLS or SD-WAN
Verdict: Pursue. The network path for every on-premises source. Direct Connect from the head office and SAP, site-to-site VPN from plants without a leased line, and one Transit Gateway per region joining both to the dev, pre-prod and prod VPCs. Nothing about Databricks changes this: it only needs your data to reach S3, Kinesis or a private endpoint without the open internet.

In plain words

What it is

A data platform in the cloud is only useful if data from the factory and the head office can get to it. Most of that data does not start in AWS. Press readings come from a plant historian on the shop floor, orders and stock come from an SAP system in a data centre, and quality results come from a lab system on a local server. These three services are the private roads between those places and your AWS network.

AWS Direct Connect is a leased physical line. A telecom partner runs a fibre circuit from your building or data centre to an AWS Direct Connect location, a colocation site where AWS has routers, and you get a private, steady connection with a known speed. AWS Site-to-Site VPN is the cheaper road: two encrypted IPsec tunnels (IPsec is the standard for encrypting network traffic between two routers) that run over the ordinary internet from a router at the plant to AWS. It can be ready in a day but its speed depends on the internet that day.

AWS Transit Gateway is the roundabout where all these roads meet. Without it, every VPC (Virtual Private Cloud, your own private network inside AWS) would need its own connection to every plant. With it, each VPC and each line or tunnel attaches once to the hub, and route tables on the hub decide who may reach whom.

A concrete example: a tyre maker has a head office with SAP S/4HANA, three plants in India on the company's own MPLS network (a private network sold by a telecom carrier), and two plants in Europe with only internet links. The head office gets Direct Connect into the Mumbai region. The Indian plants reach AWS through the head office. The European plants each get a site-to-site VPN to the Mumbai Transit Gateway. Every sensor reading, SAP extract and supplier file then travels privately, without any public IP address on the AWS side.

Direct Connect is a private toll road you lease, the VPN is an armoured van on public roads, and the Transit Gateway is the one roundabout every road and every building connects to.

Big picture

Where it fits

SourcesMove inStreamStoreProcessServeUseAcross everything: catalog · security · orchestration · observability and cost · infrastructure as codegets data fromhands data toPlants, head office and SAP data centreCompany MPLS or SD-WANIoT GreengrassDirect Connect, VPN, Transit GatewaySecurity, identity and networkVPC, endpoints and NATAWS DMSKinesis Data Streams

The lit lane is where Direct Connect, VPN, Transit Gateway works. Boxes on the left feed it; boxes on the right use what it produces.

These services sit under the ingestion lane, not in it. Every mover in the reference platform needs a path from a source to AWS: DMS reads databases at the head office, AppFlow or a certified extractor reads SAP, SiteWise Edge on Greengrass sends plant data to IoT Core and Kinesis, and Transfer Family takes supplier files. The movers that run inside your VPC reach on-premises systems over Direct Connect or VPN through the Transit Gateway, and the plant gateways send data the same way to VPC endpoints for Kinesis and S3.

The reference design is: Direct Connect from the head office, the Indian plants on the company network behind it, the European plants on site-to-site VPN to the Mumbai Transit Gateway, and one Transit Gateway per region with dev, pre-prod and prod VPCs attached. The Transit Gateway in the recovery region (Hyderabad) is peered with Mumbai, the Direct Connect gateway reaches both regions, and the European VPN tunnels to Hyderabad are configured but idle until a disaster.

Databricks itself does not need any of this to talk to its control plane; that uses PrivateLink. What these roads carry is the source data, the traffic from plant historians and SAP, and the engineers' own access from the office to the workspace front end over private connections. If a job on classic compute must reach a private system on premises, its VPC routes to the Transit Gateway; serverless compute reaches private systems only through the private connectivity options Databricks offers, so check them before you rely on it.

How it works

The ideas you need

IdeaWhat it means
Direct Connect connectionA physical port at an AWS Direct Connect location. Dedicated connections come at fixed speeds such as 1, 10, 100 or 400 Gbps; hosted connections are smaller slices sold by a partner.
Virtual interface (VIF)A logical channel on the line. A private VIF reaches VPCs, a transit VIF reaches a Transit Gateway, a public VIF reaches public AWS endpoints.
Direct Connect gatewayA global object that lets one line reach Transit Gateways or VPCs in several regions, which is how Mumbai and Hyderabad share the same line.
Customer gatewayThe record in AWS that describes your router at the plant: its public IP address and BGP number.
BGPBorder Gateway Protocol, how routers tell each other which networks sit behind them. Dynamic routing with BGP lets traffic fail over between tunnels by itself.
Transit Gateway attachmentOne connection into the hub: a VPC, a VPN, a Direct Connect gateway, or a peering link to another region's hub.
Transit Gateway route tableDecides where traffic from each attachment may go. Separate tables keep dev from reaching prod while both reach the plants.
ECMPEqual-cost multipath: spreading traffic over several VPN tunnels at once to add up their bandwidth. It needs BGP, not static routes.
VPN ConcentratorA newer option that puts up to 100 small sites, each under about 100 Mbps, behind one Transit Gateway attachment.

Choosing

When to use it, and when not

Use it when

  • Use Direct Connect when a site sends steady or large volumes, such as SAP extracts, database change capture or a historian backfill, and you need predictable speed and latency.
  • Use site-to-site VPN for plants and small offices where a leased line is too slow to order or too costly, and as the backup path behind every Direct Connect line.
  • Use a Transit Gateway as soon as you have more than two or three VPCs, or more than one site, so every network attaches once and routing lives in one place.
  • Use a Direct Connect gateway with Transit Gateway peering when a recovery region must be reachable over the same lines without new circuits.
  • Use the VPN Concentrator when you have dozens of small sites, such as warehouses or dealer workshops, each needing a little bandwidth.

Avoid it when

  • Avoid all of it when the source is a SaaS product or a cloud service; AppFlow, PrivateLink or the vendor's own private endpoint is the path, not a tunnel to an office.
  • Avoid a single Direct Connect line without a second line or a VPN backup; one cut fibre stops every ingestion job.
  • Avoid VPC peering meshes once you pass a handful of VPCs; each pair needs its own link and the routes become hard to reason about.
  • Avoid building a Transit Gateway per team; one per region, owned by the network account and shared, is easier to secure and cheaper.

Production grade

What goes wrong in production

ProblemWhat breaksWhat to do
Overlapping IP rangesA plant and a VPC both use 10.0.0.0/16, so routes clash and traffic goes to the wrong place or nowhere.Agree an IP address plan before building any VPC; reserve ranges per region and environment; use the Transit Gateway or a private NAT only as a last resort.
VPN tunnel ceilingA standard tunnel carries up to about 1.25 Gbps; a historian backfill saturates it and live sensor data lags.Use BGP with ECMP over several tunnels on a Transit Gateway, large bandwidth tunnels where supported, or Direct Connect for heavy sites; schedule backfills at night.
Single point of failureOne Direct Connect line or one plant router fails and the site goes dark for hours.Two lines at two Direct Connect locations for critical sites, or a VPN as backup with BGP so failover is automatic; test failover every quarter.
MTU and fragmentationLarge packets fail silently over VPN, which allows an MTU of 1446 bytes, and some transfers hang.Set MSS clamping on the plant router as AWS's customer gateway guidance describes; test with real file sizes, not ping.
Prod reachable from devOne flat Transit Gateway route table lets a dev notebook reach the prod database at the head office.Separate route tables per environment; plants and on-premises routes propagate only where needed; review routes in code, not in the console.
Unexpected data processing chargesEvery gigabyte from a VPC through the Transit Gateway is billed, including chatty traffic between VPCs.Keep S3 and Kinesis traffic on VPC endpoints inside the VPC; avoid routing bulk data between VPCs through the hub; watch the Transit Gateway line in the cost report.

Alternatives

What to compare it with

AlternativePick the alternative whenPick Direct Connect, VPN, Transit Gateway when
AWS Cloud WANYou have many regions and sites and want a managed global network with central policy instead of peering Transit Gateways by hand.You run one or two regions; a Transit Gateway per region with peering is simpler and well understood.
VPC peeringTwo or three VPCs need to talk and nothing else; there is no hourly hub charge.You have plants, several environments or a recovery region; peering does not pass traffic onwards and becomes a mesh.
AWS PrivateLinkYou need to reach one service, such as Databricks or a partner API, privately without joining whole networks.You need two-way routing between whole sites and VPCs; PrivateLink exposes a service, not a network.
SD-WAN appliance on EC2 with Transit Gateway ConnectThe company already runs an SD-WAN fabric and wants AWS as one more site in it.You have no SD-WAN team; plain VPN and Direct Connect are fewer moving parts.
Azure ExpressRoute and Virtual WANThe platform runs on Azure Databricks; the same design uses ExpressRoute for the line and Virtual WAN as the hub.The data and Databricks workspaces are on AWS.

Cost

How you pay

Direct Connect bills per port hour for the connection speed and per gigabyte of data transfer out of AWS; data into AWS over Direct Connect is free, and the telecom partner bills separately for the circuit. Site-to-Site VPN bills per connection hour and normal data transfer out. Transit Gateway bills per attachment hour and per gigabyte of data sent into the hub from VPCs, VPNs and Direct Connect. Peering traffic between regions adds inter-region transfer on the sending side. Ingestion into AWS is mostly inbound, which keeps the line cheap; the hub's per-gigabyte charge is what grows.

What drives the bill

  • Port speed and number of Direct Connect connections, plus the partner's circuit charges outside AWS.
  • Number of Transit Gateway attachments: every VPC, VPN and peering link pays by the hour in every environment.
  • Gigabytes processed by the Transit Gateway, especially traffic between VPCs and from plants.
  • Data going out of AWS to sites, such as reports pushed to plants or replication back on premises.
  • Inter-region transfer to the recovery region, if data is replicated over the peered hubs rather than by S3 replication.

Cost traps

  • Routing S3 traffic from a VPC through the Transit Gateway to a central endpoint VPC pays hub processing on every gigabyte; local gateway endpoints for S3 are free.
  • Idle VPN connections and attachments in the recovery region bill every hour, so count them in the standby cost.
  • A hosted connection sized for the backfill and never resized wastes money every month after go-live.

Prices change; check the AWS pricing page for your region before deciding.

Security and governance

Keeping it safe

  • Site-to-Site VPN encrypts with IPsec. Direct Connect is private but not encrypted by default; use MACsec on supported dedicated connections, or run a VPN over the Direct Connect line, when policy requires encryption in transit on the wire.
  • All traffic still lands in VPCs with no public IPs; security groups and network ACLs decide what each subnet accepts from on premises, so a plant can reach the Kinesis endpoint but not the database subnet.
  • Own the Transit Gateway in a central network account and share it with Resource Access Manager; environment accounts attach but cannot change hub routes.
  • Use separate Transit Gateway route tables to keep dev, pre-prod and prod apart; prod should not be reachable from dev even through on-premises paths.
  • Turn on VPC Flow Logs and Transit Gateway Flow Logs to the log archive account, and record every network change in CloudTrail; network changes go through Terraform pull requests.
  • Personal data in transit follows the same tokenise-at-the-door rule as the rest of the platform; the network makes the path private but does not decide who may see the values.

Operating it

What to watch, and the limits

WatchWhy
ConnectionState (Direct Connect)Whether the physical line is up; alarm on any change to down.
VirtualInterfaceBpsEgress and BpsIngressThroughput per virtual interface; alarm near the port speed during backfills.
TunnelState (VPN)Each of the two tunnels per VPN connection; alarm when either is down, not only both.
TunnelDataIn and TunnelDataOutBytes over each tunnel; shows whether traffic really spreads across tunnels with ECMP.
BytesIn, BytesOut and PacketDropCountNoRoute (Transit Gateway)Traffic per attachment and packets dropped because no route matched, often the first sign of a routing mistake.
Landing freshness per sourceThe platform's own signal: when files or events from a plant stop arriving, check the network first.
LimitValue and notes
VPN tunnel bandwidthUp to 1.25 Gbps per standard tunnel; up to 5 Gbps per large bandwidth tunnel
VPN MTU1446 bytes; no jumbo frames, no path MTU discovery
Transit Gateway attachments5,000 per Transit Gateway by default
Transit Gateway VPC attachment bandwidthUp to 100 Gbps each direction per Availability Zone
Transit Gateway routes10,000 combined routes across all route tables of one Transit Gateway
Direct Connect gateway reachUp to 6 Transit Gateways per Direct Connect gateway

Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.

Decisions

Situations and verdicts

SituationChoiceVerdict
Head office with SAP and DMS sourcesTwo Direct Connect connections, or one plus a VPN backup, on a transit VIFPursue: steady, predictable path for change capture and SAP extracts; the backup keeps landing filling when one line fails.
European plants with only internet linksSite-to-site VPN with BGP to the Mumbai Transit GatewayPursue: ready in days; size sensor streaming to the tunnel and pull the historian in batches.
Recovery regionPeered Transit Gateway in Hyderabad, Direct Connect gateway reaching both regions, VPN tunnels configured but idlePursue: recovery does not wait on a new circuit.
Five regions and many sites worldwideAWS Cloud WANPursue with care: central policy pays off at that scale; for one or two regions, Transit Gateway peering is simpler.

FAQ

Questions people ask

Do I need Direct Connect to use Databricks on AWS?

No. Databricks and your S3 data are both in AWS and talk over AWS's network and PrivateLink. You need Direct Connect or VPN only to bring data from your own sites, or to let staff reach the workspace privately from the office.

Is Direct Connect encrypted?

Not by default. It is a private line, not an encrypted one. Use MACsec on supported dedicated connections, or run an IPsec VPN over it, and keep TLS on every application connection anyway.

What is the difference between a Transit Gateway and Cloud WAN?

A Transit Gateway is a regional hub you peer with other regions yourself. Cloud WAN is a managed global network built from such hubs, with one policy document for segments and routing. Small footprints suit Transit Gateway; many regions and sites suit Cloud WAN.

How fast can a site-to-site VPN go?

A standard tunnel carries up to about 1.25 Gbps and a large bandwidth tunnel up to about 5 Gbps. On a Transit Gateway with BGP you can spread traffic over several tunnels with ECMP. Real speed depends on the internet path, so test with production-sized transfers.

Sources

Checked against

Checked 2026-11-11.

Related

Where it connects

AWS and Databricks reference platform · Ingestion on the data platform · VPC, VPC endpoints and NAT gateway · AWS PrivateLink · AWS IoT SiteWise and SiteWise Edge · All AWS data services

← Step Functions · IAM and STS →