Skip to content
Healthcare cloud migration · facts checked 13 November 2026

Healthcare cloud migration services that account for every message

Moving an EHR, an interface engine and years of clinical data to AWS or Azure is mostly a network, data and partner problem, not a server problem. This page sets out how we plan it: the partner register, static addresses and VPNs, MLLP over TLS, a counted cutover, record-level validation and a disaster-recovery site your partners can actually reach.

Server racks and network cabling in a cloud data centre
At a glance
  • Partner register firstEvery feed, IP allow-list and VPN listed before a server moves
  • Counted cutoverMessages in, out and acknowledged, reconciled per channel
  • Record-level checksRow counts, checksums and document hashes, old against new
  • DR partners can reachRecovery addresses on allow-lists before the day you need them

What healthcare cloud migration services cover

Healthcare cloud migration services move clinical systems, usually an EHR, its database and documents, an HL7 interface engine and file or API feeds, from a server room or old host to AWS or Azure without losing data or messages. The work that decides success is specific to healthcare: listing every trading partner and how it connects, re-creating VPNs and IP allow-lists, keeping PHI in services covered by the cloud provider’s BAA, reconciling every message at cutover and proving the migrated records match the originals.

EHR and interface engine cloud migration

The servers are the easy part. The partners, the queues and the records are not

A clinic group or hospital rarely moves to the cloud because it wants to. The server under the EHR is out of support (Windows Server 2016 loses extended support on 12 January 2027), a data centre contract is ending, or the one person who understood the interface box has left. Copying virtual machines to AWS or Azure takes days. What takes weeks is everything attached to them: a laboratory that only accepts traffic from one IP address, a radiology partner on a VPN that was set up in 2017, a payer SFTP drop, a pharmacy feed that holds a TCP connection open for months.

Generic migration methods treat a workload as compute, storage and a database. A clinical estate adds three things they miss. Messages are in flight at every moment and a lost result is a patient-safety event, not a retry. Patient records span a database, a document store and encryption keys that must move together from the same point in time. And the cloud provider becomes a business associate under HIPAA, so the services you pick and the logs you ship are a contractual question as well as a technical one.

We do this work as part of our cloud infrastructure and interface engine support practices. If your main problem is an old Mirth Connect release rather than the move itself, read our Mirth Connect migration guide first: it covers versions, extensions and the replay-and-diff harness. This page covers the cloud side: networks, partners, data and recovery.

What actually moves, and where it usually lands

Write the target for every component before choosing tools. The right answer differs by component: an interface engine keeps state on disk and is easiest to move as a virtual machine, while an EHR database usually benefits from a managed service.

ComponentUsual AWS targetUsual Azure targetHow it movesMain risk
EHR web and application tier (for example OpenEMR on PHP and Apache)EC2, or containers on ECS or EKSVirtual machines or Azure Kubernetes ServiceRebuild from a clean release plus your configuration; avoid copying a hand-patched serverLocal changes to core files that nobody recorded
EHR database (MySQL or MariaDB)Amazon RDS for MySQL or MariaDB, or Aurora MySQLAzure Database for MySQL Flexible Server (the MariaDB service was retired on 19 September 2025)Dump and restore for small databases; replication then switch for large onesTime zone, character set and SQL mode differences that change data silently
Documents and scanned filesAmazon EFS or S3Azure Files or Blob StorageBulk copy (AWS DataSync, azcopy) then a final deltaFiles and database rows copied at different moments
Interface engineEC2 with its own databaseVirtual machine with its own databaseFresh install with exported channels, or block replication of the whole serverPartner IP allow-lists, VPNs and messages in flight
SFTP drops and batch filesAWS Transfer Family or an SFTP server on EC2SFTP-enabled Blob Storage or a virtual machineRecreate users and keys; partners repointPartners pinning your old host key or IP
FHIR serverAWS HealthLake or a self-hosted serverAzure Health Data Services FHIR serviceBulk $export then $import, or replay through the APIAzure API for FHIR was retired on 30 September 2026; anything still pointing at it needs moving now

Physical transfer devices are less of an option than they were: AWS closed Snowball Edge to new customers on 7 November 2025 and points them to DataSync. For most clinic and small-hospital document stores, a few terabytes over a VPN or internet link with DataSync is enough; plan the bandwidth and the window.

Step 1: build the partner connectivity register

The most useful document in a healthcare cloud migration is a spreadsheet, not a diagram. One row per connection, whether or not anyone thinks it is still used. Pull it from the engine’s channel list, the firewall rule set, the VPN configuration and a week of connection logs, then confirm each row with the partner. The columns we use:

Partner and feedDirection and protocolPath todayWhat they allow-listTheir lead timeTLS and ACKPeak volume
Reference lab, orders and resultsOut: ORM over MLLP; in: ORU over MLLPIPsec VPN, policy-based on their sideOur tunnel peer IP and one inside hostChange ticket, 15 working daysInside VPN, no TLS; original-mode ACK2,000 results per hour at 7 a.m.
Hospital ADT feedIn: ADT A01–A08 over MLLPInternet, their source IP on our firewallNothing; we allow-list themNoneTLS 1.2 via proxy; ACK after commitSteady, 300 per hour
ClearinghouseOut: X12 837 batch; in: 835 and 999SFTP, key authenticationOur egress IP and our public keyPortal request, 5 daysSSH; 999 acknowledgment per batchNightly batch

The rows are illustrative. Add the partner’s after-hours contact, whether their test environment exists, and whether they can send to two addresses at once. Partners with the longest lead time set the date of your migration, not your cloud build.

Step 2: decide addresses and tunnels before building anything

  • Outbound: one fixed egress address per environment. Partners allow-list the address your traffic comes from. In AWS that means private subnets behind a NAT gateway with an Elastic IP; in Azure, a NAT gateway with a static public IP or prefix. Do not let instances send directly from their own public IPs, which change when they are replaced.
  • Inbound: addresses that survive rebuilds. An AWS Network Load Balancer can hold one Elastic IP per Availability Zone, which suits MLLP listeners because partners connect to fixed IPs. Give partners every address you will use, including the recovery region, in the first change request.
  • VPNs: one per partner, with traffic selectors agreed in writing. AWS Site-to-Site VPN connections come with two tunnels; ask the partner to configure both. Older partner firewalls often run policy-based VPNs with several encryption domains, which is the most common cause of a tunnel that connects but passes no traffic. Agree the exact local and remote subnets first.
  • Overlapping address ranges. If the partner’s network and your new VPC both use 10.0.0.0/16, one side has to translate. Choose a VPC range that no partner uses before the first subnet exists; renumbering later means another round of partner changes.
  • DNS. Where partners connect by name, cut the record’s TTL to five minutes at least a week ahead. Where they connect by IP, which is common for MLLP, the change request is the cutover.

Step 3: MLLP over TLS, or MLLP inside a tunnel

MLLP, the framing HL7 v2 uses over TCP, has no encryption of its own. Across the internet it needs either an IPsec VPN around it or TLS on top of it. Inside a VPC, many teams still encrypt it so that PHI never crosses a network in clear text. There are four workable patterns:

  1. Keep it inside the partner VPN. Simplest when the VPN already exists. The tunnel is the encryption.
  2. TLS on the engine’s own listener. Possible where your engine or its extensions support it; check this before moving off a licensed product.
  3. A TLS proxy beside the engine. stunnel or HAProxy terminates TLS and forwards plain MLLP to the engine on localhost. Engine-independent and easy to test.
  4. TLS on the load balancer. An NLB TLS listener terminates TLS with a certificate from AWS Certificate Manager. Traffic between the load balancer and the engine is then plain TCP inside the VPC unless you re-encrypt it, so decide whether your security policy accepts that.

A minimal stunnel pair for pattern 3, one inbound listener that requires partner certificates and one outbound connection that checks the partner’s name:

; /etc/stunnel/mllp.conf
sslVersionMin = TLSv1.2

[adt-in-from-hospital]
accept     = 0.0.0.0:6671
connect    = 127.0.0.1:6661
cert       = /etc/stunnel/engine.pem
key        = /etc/stunnel/engine.key
CAfile     = /etc/stunnel/partners-ca.pem
verifyChain = yes

[orders-out-to-lab]
client     = yes
accept     = 127.0.0.1:7001
connect    = mllp.lab.example:6700
CAfile     = /etc/ssl/certs/ca-certificates.crt
verifyChain = yes
checkHost  = mllp.lab.example

The engine’s outbound channel then points at 127.0.0.1:7001. Put every certificate expiry date in the partner register; an expired certificate stops a feed as surely as a deleted firewall rule.

Step 4: a cutover in which every message is accounted for

“Zero message loss” is only meaningful if you can show it. Define it as: every message a sender transmitted during the window was either acknowledged by the old engine and delivered from it, or acknowledged by the new engine and delivered from it, and the counts per channel match what the senders say they sent. Four rules make that achievable.

  • Acknowledge only after the message is stored. Check each inbound channel: if the engine sends an ACK as soon as bytes arrive and crashes before writing them to its database, the sender believes the message is delivered. Set the ACK to follow a durable write.
  • Let senders hold the queue. Most HL7 senders queue and retry when a connection drops. Confirm this per partner in the register; for the few that do not, agree a short hold window or have them pause sending.
  • Drain before you switch. Stop the inbound listeners on the old engine, then wait until every outbound queue on it reads zero, then record the last control ID (MSH-10) delivered on each channel.
  • Expect duplicates and handle them. A sender that never saw an ACK will resend. Receivers that cannot tolerate duplicates need a check on sending facility plus MSH-10.

The sequence we use on the night, with a named person against each line:

  1. Freeze channel and configuration changes on both engines.
  2. Confirm the new engine’s outbound connections from the new egress address with a test message flagged as non-production (MSH-11 = T) where the partner accepts it.
  3. Stop inbound listeners on the old engine. Senders start queuing.
  4. Wait for outbound queues on the old engine to reach zero. Export per-channel received, sent, filtered and errored counts.
  5. Switch the route: VPN tunnel, DNS record, NAT rule or partner-side IP change.
  6. Start listeners on the new engine. Watch the first connection, the first message and the first ACK on each feed.
  7. Reconcile: old counts plus new counts against each sender’s own log for the window. Investigate every gap before declaring the cutover done.

Run the switch partner by partner if you can. A parallel run before the switch, with the new engine receiving copies of live traffic, is the best way to find configuration differences; our Mirth migration guide describes the harness and the precautions for pollers.

Step 5: move PHI and prove it arrived intact

Take OpenEMR as the worked example, because its parts are typical of a self-hosted EHR: a MySQL or MariaDB database, a site directory holding documents and other files, and encryption keys kept partly in the database and partly on disk. The database and the site directory must be copied from the same moment, or encrypted fields and documents will not open after the move. OpenEMR’s own AWS packaging makes the same point about taking a database snapshot and a file backup together.

Choosing the copy method

  • Dump and restore with mysqldump --single-transaction or mydumper suits databases that restore inside your downtime window. Time a full rehearsal; restore time, not dump time, is usually the limit.
  • Replication then switch suits larger databases: restore a snapshot, replicate changes from the old server with binary-log replication or AWS DMS, then stop writes and switch. Downtime becomes minutes.
  • Files go in a bulk copy days ahead, then a final delta after writes stop.

Validation that a security lead will accept

CheckHowWhat it catches
Row counts per tableCount every table on both sides at the same frozen point; for OpenEMR start with patient_data, form_encounter, forms, documents, billing, prescriptions, lists and logMissing tables, failed batches, truncated imports
Table checksumsCHECKSUM TABLE on both sides, or ordered hashes of primary key and content per tableSilent changes from character-set or SQL-mode conversion
File manifestSHA-256 of every file under the site directory, before and after, compared by pathMissed files, partial copies, permission failures
Time and text spot checksAppointments around a daylight-saving change; names with accents and non-Latin scriptsServer time-zone differences that shift TIMESTAMP columns; mojibake
Clinical walk-throughA clinician opens ten agreed patients on old and new: demographics, last encounter, medications, a scanned document, an encrypted fieldAnything the counts cannot see, such as a broken document path
Audit trail continuityLast audit entry before the freeze present on the new side; new entries written after go-liveA gap in the record of who looked at what

Most of these checks need no human to read patient data: counts, checksums and hashes compare numbers. Keep it that way. The fewer people who see PHI during a migration, the shorter your access log and your list of business associates.

Step 6: BAA scope is an engineering decision too

This is not legal advice, and the compliance programme around it is SemperWise’s work, not ours. But several BAA questions are settled by architecture, so they belong in the migration plan:

  • The cloud provider’s BAA. On AWS you accept it in AWS Artifact, and AWS says PHI should be processed, stored and transmitted only in services on its HIPAA-eligible list. Microsoft offers its BAA through the Product Terms by default to covered entities and business associates, for the services listed as in scope for HIPAA. Check every service in your design against the current list, not last year’s.
  • Encryption does not remove the obligation. HHS guidance on cloud computing from October 2016 says a provider that stores only encrypted ePHI and holds no key is still a business associate. Encrypt anyway; it changes breach outcomes.
  • Logs are the usual leak. Interface engines log message content. Ship those logs to a monitoring SaaS without a BAA and PHI has left your covered boundary. Mask message bodies at the engine or keep message logs in covered services.
  • Support tickets and attachments. A sample HL7 message pasted into a vendor ticket is a disclosure. Agree de-identified test messages for every vendor conversation.
  • The migration team. Anyone with access to PHI during the move, including us, needs to be covered by an agreement before access is granted. Design the work so that most of it needs no PHI access at all.

Our HIPAA on AWS guide maps the Security Rule’s safeguards to AWS controls in more detail.

Step 7: disaster recovery that partners can reach

Recovery targets are a business decision; agree them per system before choosing a pattern.

PatternTypical RTOTypical RPOSuitsPartner implication
Backup and restore to another regionHoursTime since last backupReporting, archives, small practicesPartners must change allow-lists on the day unless addresses were pre-registered
Block-level replication (AWS Elastic Disaster Recovery, Azure Site Recovery)Minutes to an hour, rehearsedSeconds to minutesInterface engines and EHR servers kept as virtual machinesRecovery egress IP and VPN peers must already be on partner allow-lists
Warm standby with database replicaMinutesSecondsEHR databases on managed servicesSame as above, plus DNS with a short TTL
Active in two Availability ZonesSeconds to minutes for a zone failureNear zeroEvery production clinical system, as a baselineNone if the load balancer keeps its addresses

AWS describes Elastic Disaster Recovery as enabling “RPOs of seconds and RTOs of minutes”. For an interface engine, the replication is rarely the slow part. The slow part is a partner whose firewall does not know your recovery address. Put recovery IPs and VPN peers in the first round of partner changes, and test a failover once a year with each partner that agrees to it.

What a healthcare cloud migration costs to run

Compute and storage are what calculators show. These lines are what surprise people in the first bill (AWS list prices from the VPN and VPC pricing pages, US East (Ohio) examples, checked 13 November 2026):

  • Site-to-Site VPN: $0.05 per connection-hour, about $36 a month for each partner tunnel, plus data transfer out.
  • NAT gateway: $0.045 per hour plus $0.045 per GB processed. Large document transfers through NAT add up quickly; use VPC endpoints for S3 traffic.
  • Public IPv4 addresses: $0.005 per hour each, in use or idle.
  • Replication during migration: AWS Transform MGN (formerly AWS Application Migration Service) is free for 2,160 hours per source server, but the staging instances and volumes it creates are billed meanwhile.
  • Double running: old and new environments both run through the parallel run and hypercare. Budget four to eight weeks of overlap.
  • Support plans and licences: a cloud support plan, Windows Server and SQL Server licences, and any commercial engine licence that is priced per server.

Once the estate is stable, our cloud cost optimization page covers rightsizing and commitments.

Illustrative plan: a clinic group moving OpenEMR and its engine to AWS

Illustrative, not client work. Twelve clinics share one OpenEMR instance on MariaDB, about 400 GB of database and 2 TB of documents on two ageing on-site servers, plus an Open Integration Engine with 18 channels and 9 partners: two labs on VPNs, a hospital ADT feed over the internet, a radiology group, a clearinghouse over SFTP and four smaller feeds. Target: AWS, two Availability Zones, RDS for MariaDB, EFS for documents, the engine on EC2, recovery in a second region.

PhaseElapsedWorkTypical effort
Discovery and partner registerWeeks 1–2Inventory, register, partner change requests sent on day 5, target design, data method chosen5–7 engineer-days
Landing zone and networkWeeks 2–4Accounts, VPC ranges checked against partners, NAT and load balancer addresses, VPNs, logging in covered services6–9 engineer-days
EHR rebuild and data rehearsalWeeks 3–6OpenEMR from a clean release plus configuration, two timed restore rehearsals, validation scripts8–12 engineer-days
Engine move and parallel runWeeks 5–8Engine on EC2, TLS proxy, channels imported, copied traffic compared daily6–10 engineer-days
CutoverWeek 9Overnight window: engine by partner wave, EHR with final delta, reconciliation4–6 engineer-days
DR build and hypercareWeeks 9–12Replication to second region, first failover test, monitoring, old servers wiped after sign-off5–8 engineer-days

Total: roughly 34–52 engineer-days over about twelve weeks. Partner change lead times stretch it more than anything technical. Rebuilding OpenEMR from a clean release also makes the next step easy: OpenEMR 8.5.0 came out on 27 October 2026, and an unmodified install can be upgraded on its own schedule after the move. Move first, upgrade second; never both in one window.

Questions to settle before you approve a migration plan

  1. Is there a partner register, and has every partner confirmed its row and lead time?
  2. Which egress and inbound addresses, including recovery, have partners been given?
  3. Where does each inbound channel send its ACK: on receipt or after a durable write?
  4. How will the cutover be reconciled, and against whose counts?
  5. Which validation checks run on the EHR data, and who signs them off?
  6. Is every service that touches PHI, including logging and monitoring, covered by the provider’s BAA?
  7. What are the agreed RTO and RPO, and when is the first failover test?
Services

Healthcare cloud migration services

Effort is shown in engineer-days for a single EHR with up to about 20 interface channels. These are typical ranges; we fix a price per phase after the discovery.

Migration discovery and partner register

The problem

You know the systems must move but not what is connected to them, who must change what, or how long partners will take.

How we do it

We build the partner connectivity register from engine exports, firewall rules, VPN configuration and connection logs, confirm it with partners, choose targets per component and write the plan. Typical effort: 5–8 engineer-days.

What you get

  • Partner connectivity register
  • Target design for AWS or Azure
  • Data copy method with timed estimate
  • Phased plan with effort per phase
How it runs

How a healthcare cloud migration runs with us

Typical for one EHR and up to about 20 interface channels. More partners add waves, not stages.

Week 1

Discovery

Read-only access or exports; the partner register started and change requests drafted for the slowest partners.

Weeks 2–4

Foundation

Landing zone, addresses, VPNs and logging in covered services, built as code.

Weeks 3–8

Rehearse

EHR rebuilt, data copies timed, validation scripts run, engine parallel run with daily comparisons.

Week 9

Cutover

Counted cutover by partner wave, EHR switch with final delta, reconciliation signed off.

Weeks 9–12

Recover and hand over

DR built and tested, monitoring in place, old hardware wiped, runbooks handed to your team or to ours.

Senior engineers firstYour first call is with someone who will do the work, not a sales script.
You own everythingCode, configuration, reports and documentation stay in your accounts and are handed over.
We run our own products15 live platforms we build and operate, so we feel production pain too.
Honest about fitIf a cheaper route or another provider suits you better, we will say so.
FAQ

Questions about healthcare cloud migration

For one EHR with up to about 20 interface channels, typically ten to fourteen weeks and 30 to 55 engineer-days. The technical work rarely sets the pace. Partner firewall and VPN change requests do, so they are sent in the first week.

Not quite, but close. With database replication and a pre-copied document store, the outage is the time to stop writes, apply the final changes, switch and check, usually under an hour. Interface feeds keep queuing at the sender during that window, so no messages are lost if acknowledgments are set up correctly.

That every message a sender transmitted during the window was acknowledged and delivered by either the old or the new engine, and the per-channel counts match the senders’ own logs. It depends on acknowledging only after a durable write, letting senders queue, draining the old engine before the switch and handling resends as duplicates.

Build the engine on EC2 with a database it owns, give it a fixed egress address through a NAT gateway and fixed inbound addresses through a Network Load Balancer, recreate partner VPNs, and add a TLS proxy for MLLP where needed. Then run it in parallel and cut over partner by partner. If you are also changing Mirth versions, do that as a separate change.

Both offer a BAA and run EHRs and interface engines well. Choose on your team’s skills, existing licences and the services you need. Note that Azure retired its MariaDB service in September 2025 and Azure API for FHIR in September 2026, so MariaDB-based EHRs and older FHIR deployments need extra planning on Azure.

No. The BAA covers the provider’s side for eligible or in-scope services only. You still need to use only those services for PHI, configure them securely and run your own compliance programme. Our sister company SemperWise handles compliance programmes; we handle the engineering.

By comparing numbers rather than reading records: row counts and checksums per table from the same frozen point, a SHA-256 manifest of every document file, targeted checks on time zones and special characters, and a short walk-through of agreed patients by a clinician. The results go into a validation report your security lead signs.

At minimum, production across two Availability Zones and tested backups in a second region. For interface engines, replicate the server and register the recovery addresses with every partner in advance, because a partner who has not allow-listed the recovery IP turns a fifteen-minute failover into a multi-day outage.

Last updated