… 10 unchanged lines …
  - For customer engagement apps the claim is a significant improvement in delta sync latency rather than a published number.
  - Removing the CSV stage is pitched as a reliability improvement as much as a speed one — fewer failure points, less resource contention, simpler retry and recovery, and no CSV-to-delta encoding or schema mismatch edge cases.
  - The same announcement restated that Export to Data Lake is deprecated and being retired, with no further extension or exception process planned.
+ - A **new SQL analytics endpoint metadata sync** is in preview alongside it (documented September 2026). Low-latency sync lands the data fast through shortcuts; the SQL analytics endpoint then has to pick it up, and the v2 metadata sync is the piece that stops that becoming the new bottleneck.
+ - During preview, SQL endpoint v2 works only with a **new** workspace connection — you select a workspace with the new metadata sync enabled when you create the Link to Fabric. It isn't retrofitted to an existing link.
+ - Change data feed (CDF) is memory- and compute-intensive and can raise latency even on low-latency sync. Microsoft now says this explicitly rather than leaving it to be discovered.
  
  ## When to use / skip
  There's nothing to decide about whether you get it — it arrives with your station. What you decide is whether to force it early. For a client already live on Fabric Link with acceptable latency, unlinking and relinking to pick it up means a full initial sync across every table, and that's a real outage window for downstream reports. Wait. For a client about to set up Fabric Link for the first time, check whether their station is enabled before committing to a go-live date, because starting on the old engine and migrating later costs you that same full resync. And if latency is the actual complaint driving the project, this is the answer and it's worth planning the relink around.
    … 6 unchanged lines …
  
  ## Gotchas
  - The headline throughput figure is a lab number with an explicit disclaimer attached. Quote it in a client meeting and you own it. Talk about the architecture change instead.
- - Unlink and relink is not a small operation. It removes the Fabric lakehouse and rebuilds from scratch.
+ - Unlink and relink is not a small operation. It removes the Fabric lakehouse and rebuilds from scratch. As of September 2026 the docs also warn that unlinking **fails outright** if Fabric items still reference the lakehouse or its SQL analytics endpoint — remove or repoint dependent artefacts first, or the relink you scheduled for a Saturday night doesn't start.
  - Rollout is by station, not by tenant or region as most people picture it, so two environments in the same organisation can be on different engines for weeks.
  - Documentation for this is currently thinner than the announcement. The detail lives in the June 2026 Power Platform blog post and the `aka.ms/lowlatencylearnmore` shortlink; the Learn pages for Link to Fabric describe the feature's behaviour but don't yet carry the rollout detail. If you need a citable prerequisite list for finance and operations, chase the Learn page rather than the blog.
  - It doesn't help Synapse Link at all. Clients who assume "the Fabric sync got faster" applies to their bring-your-own-lake profile will be disappointed.
    … 6 unchanged lines …
  
  ---
  
- *Revisit once the rollout completes globally and the throughput guidance lands on a Learn page rather than a blog post.*
+ *Revisit once the rollout completes globally, and again when SQL analytics endpoint metadata sync v2 leaves preview.*