A bot that works fine in your test server will fall over the moment it joins a large community. The gateway payload volume grows with member count, and the default cache keeps all of it in memory.
Start by disabling the caches you don't use. Most bots never need the full member or presence cache, and turning them off can cut memory usage by 80%.
Shard early. Even if you are under the required threshold, running two shards forces you to write code that does not assume a single process — which makes the real migration boring instead of scary.
Finally, treat rate limits as part of your design, not an error case. Queue outgoing messages, batch role updates, and back off gracefully so a spike never gets your token flagged.