2 ms·

>if the rollback journal is present when sqlite opens a database, it applies the rollback journal to undo the changes that the transaction made. See https://www

>if the rollback journal is present when sqlite opens a database, it applies the rollback journal to undo the changes that the transaction made. See https://www.sqlite.org/atomiccommit.html#1_deleting_the_roll https://www.sqlite.org/atomiccommit.html#1_deleting_the_roll... I don't follow. How would fsyncing the rollback journal affect the durability of the actual database? Do you actually think that the database would reapply an already committed journal whose ID in the header already indicates that the transaction was committed, when the database is already consistent? I really think you should re-review the definition of durability of a database, especially before saying the creator of SQLite is incorrect about its implementation.

int_19h 1y ago
It's specifically about fsyncing journal deletion. The problem isn't that it would reapply it if it was already used to rollback. Rather, the problem is that if you commit, and that commit has succeeded (and so your app believes that it has written the data and might e.g. perform some other actions on it), the deletion of the now-unneeded journal might not be flushed to disk in event of power loss or similar. So when you start the app again, SQLite sees said rollback journal, and - since it would be considered "hot" - applies it, effectively reverting the transaction that was supposedly already committed. FWIW I don't think it's wrong per se. The article links to a HN comment in which Richard Hipp explains why this is the default behavior, and it does make sense: https://news.ycombinator.com/item?id=45014296 https://news.ycombinator.com/item?id=45014296. At the same time, clearly, the definition of "durable" here could use some clarification.
- agwa 1y ago
  Yes, that's exactly right. Note that the comment by Richard Hipp is justifying why WAL mode is not durable by default. It's a completely reasonable explanation, and would be for DELETE mode too, yet his comment claims that DELETE mode is durable by default, which I can't reconcile with the docs.
- NewsaHackO 1y ago
  >So when you start the app again, SQLite sees said rollback journal, and - since it would be considered "hot" - applies it, effectively reverting the transaction that was supposedly already committed. Guys. The journal would not be a hot journal though, as the hot journal selection only applies if the database is in a inconsistent state. Otherwise, the database knows from the ID of the journal not to reapply an already applied rollback journal. The process you are talking about ONLY happens when the journal database has been corrupted state, and it has to try and file a file to help recover the database.
  - agwa 1y ago
    The docs list 5 conditions that all must be satisfied for the journal to be considered hot: https://www.sqlite.org/atomiccommit.html#_hot_rollback_journals https://www.sqlite.org/atomiccommit.html#_hot_rollback_journ... I believe they would all be satisfied. I don't see any mention of checking IDs. Not saying you're wrong - I think the docs could very well be wrong - but could you provide a citation for that behavior?
    - NewsaHackO 1y ago
      Please read the entire document instead of just picking out sections; you will then be able to see where your misconceptions are occurring. You have attempted to make this same point three times, so I will say it for a third time; that section is about CORRUPTED databases, not database that are consistent after fsync.
      - agwa 1y ago
        I read the whole document. It doesn't mention IDs anywhere. If you're not going to provide citations for your claims, yet criticize me for "picking out sections" when I provide citations, then continuing this conversation won't be productive.
      - aktiur 1y ago
        The text in this document also directly contradicts what you're saying. Put another way: the presence of a hot journal is how SQLite determines the database might be corrupted. https://sqlite.org/lockingv3.html#hot_journals https://sqlite.org/lockingv3.html#hot_journals
agwa 1y ago
I'm pretty sure I understand what durability means; the definition is not hard - https://en.wikipedia.org/wiki/Durability_(database_systems) https://en.wikipedia.org/wiki/Durability_(database_systems) It's possible I've misunderstood how DELETE mode works. But here's the thing - I shouldn't have to understand how DELETE mode works to know what SQLite setting I need to use to get durability. Unfortunately, the SQLite docs don't clearly say what guarantees each setting provides - instead they talk about about what SQLite does when you choose the setting, leaving the reader to try to figure out if those actions provide durability. And the docs really make it seem like you need synchronous=EXTRA in DELETE mode to get durability, for the reasons explained above. This is a docs problem; I'm not saying SQLite is buggy.
- NewsaHackO 1y ago
  This may just be a expectations difference then. I would fully expect a developer to read the docs and know how a settings works to know what guarantees it has.

int_19h 1y ago

It's specifically about fsyncing journal deletion. The problem isn't that it would reapply it if it was already used to rollback. Rather, the problem is that if you commit, and that commit has succeeded (and so your app believes that it has written the data and might e.g. perform some other actions on it), the deletion of the now-unneeded journal might not be flushed to disk in event of power loss or similar. So when you start the app again, SQLite sees said rollback journal, and - since it would be considered "hot" - applies it, effectively reverting the transaction that was supposedly already committed. FWIW I don't think it's wrong per se. The article links to a HN comment in which Richard Hipp explains why this is the default behavior, and it does make sense: https://news.ycombinator.com/item?id=45014296 https://news.ycombinator.com/item?id=45014296. At the same time, clearly, the definition of "durable" here could use some clarification.

agwa 1y ago
Yes, that's exactly right. Note that the comment by Richard Hipp is justifying why WAL mode is not durable by default. It's a completely reasonable explanation, and would be for DELETE mode too, yet his comment claims that DELETE mode is durable by default, which I can't reconcile with the docs.
NewsaHackO 1y ago
>So when you start the app again, SQLite sees said rollback journal, and - since it would be considered "hot" - applies it, effectively reverting the transaction that was supposedly already committed. Guys. The journal would not be a hot journal though, as the hot journal selection only applies if the database is in a inconsistent state. Otherwise, the database knows from the ID of the journal not to reapply an already applied rollback journal. The process you are talking about ONLY happens when the journal database has been corrupted state, and it has to try and file a file to help recover the database.
- agwa 1y ago
  The docs list 5 conditions that all must be satisfied for the journal to be considered hot: https://www.sqlite.org/atomiccommit.html#_hot_rollback_journals https://www.sqlite.org/atomiccommit.html#_hot_rollback_journ... I believe they would all be satisfied. I don't see any mention of checking IDs. Not saying you're wrong - I think the docs could very well be wrong - but could you provide a citation for that behavior?
  - NewsaHackO 1y ago
    Please read the entire document instead of just picking out sections; you will then be able to see where your misconceptions are occurring. You have attempted to make this same point three times, so I will say it for a third time; that section is about CORRUPTED databases, not database that are consistent after fsync.
    - agwa 1y ago
      I read the whole document. It doesn't mention IDs anywhere. If you're not going to provide citations for your claims, yet criticize me for "picking out sections" when I provide citations, then continuing this conversation won't be productive.
    - aktiur 1y ago
      The text in this document also directly contradicts what you're saying. Put another way: the presence of a hot journal is how SQLite determines the database might be corrupted. https://sqlite.org/lockingv3.html#hot_journals https://sqlite.org/lockingv3.html#hot_journals

agwa 1y ago

I'm pretty sure I understand what durability means; the definition is not hard - https://en.wikipedia.org/wiki/Durability_(database_systems) https://en.wikipedia.org/wiki/Durability_(database_systems) It's possible I've misunderstood how DELETE mode works. But here's the thing - I shouldn't have to understand how DELETE mode works to know what SQLite setting I need to use to get durability. Unfortunately, the SQLite docs don't clearly say what guarantees each setting provides - instead they talk about about what SQLite does when you choose the setting, leaving the reader to try to figure out if those actions provide durability. And the docs really make it seem like you need synchronous=EXTRA in DELETE mode to get durability, for the reasons explained above. This is a docs problem; I'm not saying SQLite is buggy.

NewsaHackO 1y ago
This may just be a expectations difference then. I would fully expect a developer to read the docs and know how a settings works to know what guarantees it has.