The strangler fig
When a split really is justified, how you do it matters as much as the decision to do it.
Use the pattern with the best track record, sometimes called the strangler fig. Put a routing layer in front of your monolith, carve off one capability at a time into a new service, send that traffic to it, and let the monolith shrink until what remains is gone or small enough not to matter.
Notice the property that makes it work: at every single point during the migration you have a working system in production.
Choose your first extraction strategically. Good candidates sit at a natural seam, with few callers, a clear data boundary, and ideally a real pain like needing to scale differently. Sending notifications, processing images and building search indexes turn up early in most migrations for exactly those reasons.
Avoid the tangled heart of your domain, like orders, as a first candidate. You would be learning to operate distributed systems on your highest-risk component.
The data is the hard part
Plan the data explicitly, because it is the hard part. A common sequence has your new service write to its own store while the monolith keeps writing to the old tables, with change events keeping the two aligned.
Move the reads next, first behind a shadow that compares the new path against the old, then for real. Only once the monolith stops touching those tables do you delete them.
Skip steps there and you land in the distributed monolith from the last lesson.
Set expectations with your leadership before you start. This takes quarters, it will slow feature work while it runs, and the finish line is not zero monolith. Plenty of successful companies run a smaller calmer monolith surrounded by a handful of services indefinitely, and that is a destination rather than a failure.
Worked example
Marco leads platform at a marketplace with a 9-year-old PHP monolith and 80 engineers stuck in a one-deploy-per-day queue. Instead of a rewrite, his team puts Envoy in front and picks search as the first extraction: it has a clean seam, its Elasticsearch cluster already lives outside the monolith, and its traffic peaks 5x higher than checkout's. They ship the search service in six weeks, run it in shadow mode for two more comparing responses (0.3 percent mismatch, all timezone bugs in the old code), then flip the route. Over 18 months, four more services follow: notifications, payments, image processing, and seller analytics. The monolith is still there in year two, one third its old size, deploying in 4 minutes, and nobody is trying to kill the rest of it.