Industries / Technology / Software and SaaS / Cloud and data centers / Edge computing
Software
This is the software that installs, runs and updates programs on computers outside big data centers, such as a shop's back-room server, a factory gateway or a delivery network's city site, and keeps their data in step with the cloud when the network comes and goes.
6
Milestones achieved, from 2009 to 2025.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2009
Nearby cloudlets proposed
Four researchers argued in IEEE Pervasive Computing that a phone should hand heavy work to a "cloudlet", a trusted computer or small cluster reached over a local wireless network, not to a distant data center whose delay makes interactive apps sluggish. The cloudlet would load the needed software on the spot as a virtual machine (a whole computer, operating system included, packed into a file), and their prototype showed that this was feasible.
LessonThe founding argument was about delay: some work cannot wait for a round trip to a faraway building.
Credit
Mahadev Satyanarayanan, Paramvir Bahl, Ramón Cáceres and Nigel Davies
2012
Fog computing gets a name
Flavio Bonomi and three colleagues presented "Fog Computing and Its Role in the Internet of Things" at a workshop of the SIGCOMM conference in Helsinki on 17 August 2012; fog means placing computing, storage and control between devices and the cloud. Others took up the word: ARM, Cisco, Dell, Intel, Microsoft and Princeton University founded the OpenFog Consortium on 19 November 2015, and in June 2018 the IEEE adopted its reference architecture as standard IEEE 1934.
Credit
Flavio Bonomi, Rodolfo Milito, Jiang Zhu and Sateesh Addepalli
2017
Cloud code runs offline
AWS announced Greengrass in November 2016 and made it generally available on 7 June 2017. It is software for a local device that runs the same small functions written for AWS Lambda, Amazon's service for running code on demand, passes messages between nearby devices, keeps working "even when not connected to the Internet", and syncs data back later; SiliconANGLE counted twelve launch partners, among them Intel, Qualcomm, Raspberry Pi and Samsung.
LessonAn edge runtime ties devices to its maker's cloud: AWS ended support for Greengrass version 1 on 7 October 2026, and because version 1's over-the-air updates cannot install version 2, each device had to be upgraded separately.
Credit
Amazon Web Services
2018
Functions move into delivery networks
Amazon's Lambda@Edge, generally available in July 2017, put small Node.js functions into the points of presence of AWS's delivery network, the sites that serve nearby users. Cloudflare opened Workers to all customers on 13 March 2018, running customers' code in what The Register counted as its 127 data centres. Each script is kept apart inside the V8 JavaScript engine, as the Chrome browser keeps web pages apart, so no customer needs a virtual machine that "must be allocated hundreds of megabytes of RAM".
Credit
Cloudflare (Kenton Varda); Amazon Web Services
2019
Kubernetes shrunk for small machines
Rancher Labs released K3s on 26 February 2019: a cut-down Kubernetes (software that starts, stops and moves containers, packaged programs with everything they need to run, across a group of machines) of about 40 MB that runs in less than 512 MB of memory, on x86 and Arm chips. KubeEdge, which Huawei Cloud open-sourced in November 2018 to run Kubernetes on remote edge nodes, graduated, the top maturity level of the Cloud Native Computing Foundation (CNCF), in 2024.
LessonAn outside comparison by Koziolek and Eskandani in 2023 found that K3s and k0s led their lightweight rivals on control-plane throughput only "by a small amount", so speed alone does not decide between them.
Credit
Rancher Labs (K3s); Huawei Cloud (KubeEdge)
2025
An offline sync service ends
In September 2024 MongoDB deprecated Atlas Device Sync and its Device SDKs (formerly Realm), a database on the device plus a service that kept it in step with MongoDB's cloud, and it removed them on 30 September 2025. The device database continues as an open-source project but the sync service does not, so apps had to move to another sync service or drop sync.
LessonAn offline-first app, one built to keep working without a connection, depends on the server side of its sync as much as on its local database.
Credit
MongoDB
Watch for
WebAssembly functions at scale
Fermyon says its Wasm Functions, built with Akamai and running on Akamai's network, are generally available. WebAssembly is a compact, sandboxed code format that starts quickly. Fermyon claims "proven scale of up to 75 million requests per second across edge and cloud in production" and cold starts under a millisecond (a cold start is the delay when a function is first called), without saying how the figures were measured. Akamai bought Fermyon on 1 December 2025.
Achieved whenan independent party publishes measured cold-start times and peak throughput with its method, or a named customer reports its own figures.
Claimed by
Fermyon Technologies with Akamai, November 2025
One standard for edge apps
Margo, an initiative under the Linux Foundation, published Preview Release 1 of a shared specification for how industrial edge applications are defined, deployed and monitored by fleet-management software, with a sandbox for trying it, on 13 January 2026, and plans its first general release for 2026. On 19 November 2026 its GitHub releases page listed no releases.
Achieved whena general release is published and compliance tests show applications and devices from at least two makers managed by a third maker's fleet software.
Claimed by
the Margo Initiative, January 2026
Open
01
Power cuts on small clusters
A test bench that cuts power to cheap single-board computers running each lightweight Kubernetes (K3s, k0s, MicroK8s and a KubeEdge edge node) 100 times each at random moments, and records whether the machine and its containers come back without anyone touching them, and how long that takes.
The briefCredit · unclaimed02
Hours offline, then rejoin
A test any sync engine can be run through. It cuts an edge app off from its sync server for 1, 8 and 24 hours while writes continue on both sides, restarts the device during the outage, then reconnects, and counts writes lost, writes duplicated and copies that disagree.
The briefCredit · unclaimed03
Rescue kit for Greengrass V1
A tool to run on a device still on Greengrass version 1. It reads the group settings V1 left on the device and its local copies of Lambda functions, writes matching Greengrass version 2 component recipes and a deployment, installs V2 reusing the device's existing certificate, and checks that the same messages still flow.
The briefCredit · unclaimed
Open 01 · the brief
Power cuts on small clusters
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test bench that cuts power to cheap single-board computers running each lightweight Kubernetes (K3s, k0s, MicroK8s and a KubeEdge edge node) 100 times each at random moments, and records whether the machine and its containers come back without anyone touching them, and how long that takes.
Why it is still open
Published comparisons measured speed and resource use (Koziolek and Eskandani, 2023) or simulated network outages (Yakubov and Hästbacka, 2025), not power loss. K3s maintainers warn that Raspberry Pis booting from SD cards are "highly likely to experience filesystem corruption", but give no counts.
Built with what exists
Raspberry Pi 4 or 5 boards, booting from SD cards and from USB solid-state drives; a relay board or network-switched power strip; K3s with its SQLite store and with embedded etcd; k0s; MicroK8s; KubeEdge; a script on a second computer that cuts power and polls each cluster with kubectl.
First step
Wire one K3s board on an SD card to a relay, run 20 cuts, and publish the script and the first counts.
Done looks like
A public table for each distribution, storage type and datastore, from 100 cuts each, giving how many recovered unaided, the median and worst time until all containers ran again, and how many needed the storage rewritten from scratch, published with raw logs and scripts so anyone can rerun it.
Open 02 · the brief
Hours offline, then rejoin
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test any sync engine can be run through. It cuts an edge app off from its sync server for 1, 8 and 24 hours while writes continue on both sides, restarts the device during the outage, then reconnects, and counts writes lost, writes duplicated and copies that disagree.
Why it is still open
The broadest public benchmark, offline-sync-bench, covers PowerSync, Electric, Turso, Zero and Jazz; its longest outage is 20 seconds and it has no check for duplicates. The Jepsen-style test of PowerSync covers one product, with network breaks of 0 to 10 seconds.
Built with what exists
offline-sync-bench and its existing adapters; Jepsen and LazyFS, which throws away writes not yet saved to disk, as a power cut would; Linux iptables to cut the network; Docker. AI coding agents can draft each engine's adapter against the bench's existing tests.
First step
Fork offline-sync-bench, add one 8-hour outage case with a duplicate-write check, and run it on two engines.
Done looks like
For at least four engines, 30 runs at each outage length, with published counts of lost, duplicated and diverged records (the target is zero), and the harness public for reruns.
Open 03 · the brief
Rescue kit for Greengrass V1
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool to run on a device still on Greengrass version 1. It reads the group settings V1 left on the device and its local copies of Lambda functions, writes matching Greengrass version 2 component recipes and a deployment, installs V2 reusing the device's existing certificate, and checks that the same messages still flow.
Why it is still open
AWS ended support for V1 on 7 October 2026. Its upgrade guide says to complete the steps "for each device to upgrade" and recommends "that you create a script"; no converter was found.
Built with what exists
The group.json file V1 keeps under greengrass-root/ggc/deployments/group; the Greengrass V2 nucleus or nucleus lite; AWS's legacy subscription router component, which keeps V1-style message routes working; the AWS IoT Device SDK V2; AI coding agents such as Claude Code, which rewrite Lambda handlers as plain components and keep changing the code until tests pass, the test-steered method AWS described for building its cloud emulator (see the Note "Code no person read").
First step
Write a group.json parser that lists every function, subscription and resource, and test it on files from one V1 device.
Done looks like
On three test devices with different V1 setups (Lambda functions only; local client devices with subscriptions; stream manager), the kit produces V2 deployments with no hand edits, a message-flow test passes on each, and the time per device is reported.