I used to ask about performance and people would tell me that PostgreSQL keeps small-enough tables entirely in RAM automatically, as to not stress the HDD and thus enable super-fast reading/writing of "small tables", that is, with very few rows which more than easily fit in the available RAM (or what's assigned to PG).

This made sense to me, and made me very happy as I can efficiently communicate between my parallel-running scripts without stressing my poor HDD/SSD and clogging up all the resources.

But then it struck me: How is this possible? How can PG possibly keep a table all in RAM, regardless of how small it is, without losing data integrity?

If it doesn't write to the permanent storage all the time, how can it possibly recover if the power is suddenly cut, or there's a software crash, or some other disaster happens while it's running? It doesn't add up to me the more I think about it.

The only answer I can think of is: No, it doesn't actually do this. I don't see any possible way that PG could avoid losing data if it never (or rarely) writes the data to the permanent storage.

If this is possible, but only after enabling some kind of flag for the table, or setting, I'd like to know about that flag/setting. I'd like to enable this for tables which only contain "internal communication" data of the kind which has no value besides as a "common data storage" for actively running scripts.


2 Answers 2


This is possible due to Write-Ahead Logging (WAL):

Briefly, WAL's central concept is that changes to data files (where tables and indexes reside) must be written only after those changes have been logged, that is, after log records describing the changes have been flushed to permanent storage. If we follow this procedure, we do not need to flush data pages to disk on every transaction commit, because we know that in the event of a crash we will be able to recover the database using the log: any changes that have not been applied to the data pages can be redone from the log records.

It's a central component of Postgres' architecture and turned on by default. You can configure a couple of parameters like instructed in the manual chapter WAL Configuration.

Also consider basics on Reliability in the manual.

  • But even with the WAL, it still has to actually write the data to the disk, albeit in a "log" format rather than "for real"? Is that not basically the same thing in terms of strain on the I/O? Commented Aug 30, 2020 at 4:53
  • @RhomanTawadros: of course, writes are not "kept in RAM", but writing to WAL first ensures no writes are lost, and is more efficient than writing each change individually.
    – user1822
    Commented Aug 30, 2020 at 7:20
  • 3
    The change gets written to disk only, synchronously to the WAL, and the table change itself gets written only to memory. Then your synchronous work is done! The extra cost of writing the table change to disk can be done lazily. If there's a crash before it gets persisted to disk, the crash recovery process can read from the WAL to get things back where they should be, with no data loss. The data is still written to disk, but that can all be asynchronous.
    – AMtwo
    Commented Aug 30, 2020 at 13:31

PostgreSQL keeps frequently/recently used data in shared_buffers. It doesn't specifically target small tables. Each page is treated separately. A frequently accessed page from a large table or an index is just as likely to be found in shared_buffers as one from a small table that accessed with the same frequency.

There are two ways this data is protected. Another answer already described WAL. The other mechanism is checkpoints. At checkpoint time, all the dirty data from shared_buffers is written out to disk, and finally synced. But, the copy that was in shared_buffers does not get invalidated, it is just marked as clean. So it is still there and won't need to be read again if it is needed again.

So if all the data is kept in shared_buffers, that means the page only needs to be read once, and only needs to be written once per checkpoint (which defaults to 5 minutes) in which it was dirtied (written just once, no matter how many times it was dirtied). But the changes also have to be written into the WAL stream every time it is updated.

If the table is marked as unlogged, then the changes to those tables no longer need to be added to the WAL, and at checkpoint the data is not written out, except if it is a "shutdown" checkpoint. At the end of a transaction, the commit record still does get written into WAL, but if every table changed in the transaction is unlogged, then the commit record does not need to be immediately synced to disk.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.