The sequencer decides what order transactions execute in and gives users fast soft confirmations before anything reaches the base chain. Almost every major layer-2 runs one centralised sequencer today.
That creates real exposure: it can censor or reorder transactions, it captures mev from ordering, and if it goes offline the chain stops accepting new activity. Well-designed L2s provide a forced-inclusion escape hatch through the L1, though it is slow.
Example: a sequencer outage of two hours during a sharp drop means positions on that L2 cannot be adjusted or liquidated normally. If you run leverage on an L2, know the escape-hatch delay before you need it.