My understanding is as follows:
User requests are first sent to the primary shard, then synchronized to all replica shards. After all shards complete the write operation, Elasticsearch returns success to the user.
If asynchronous translog is enabled, even if one node goes down, the other replica shards still have the correct data, ensuring that data is not lost.
If my understanding is correct, then if the number of replicas is greater than or equal to 1, asynchronous translog can be enabled in any situation to reduce disk IO load, and there is no need to worry about data loss.
Setting number_of_replicas: N doesn't guarantee that there are N replicas accepting all writes. If replicas fail then ES will still accept writes and record them only on the remaining shard copies, which could be just the sole primary. In that situation if you're using async translog mode and you lose power on the primary then the write could be acknowledged and then lost.
I understand that if async translog is enabled, data may be lost if there are no replicas set or if both the primary shard and all replicas are down. But I want to know if, in the scenario where replicas are enabled and only one node goes down at a time, can data be guaranteed not to be lost in this case?
Not really, no, that is not guaranteed. Even if only one node goes down at once, it can take quite some time (minutes, maybe hours) to rebuild all the missing replica shards. During that time, if another node goes down and you're using async translog mode then data can be lost.