Old implementation could never finish if the write is faster than
time needed to do two round-trips to primary.
For example it's enough to have a write every 200ms to make sync
never finish if it's done in embedded replica in Sydney
if the primary and the writer are in the US.
Signed-off-by: Piotr Jastrzebski <piotr@chiselstrike.com>
This commit fixes an issue where new connections for embedded replica's
with write delegation would not create a new "connection" on the server
causing weird issues with transactions and state. The fix here is to
create a new `client_id` when creating a new writer for a new connection
locally. This forces the server to use separate connections to execute
the writes.
Closes#1283
We have a user with a failing snapshot, but we don't know why.
We know rename returns EINVAL, which could happen for a variety
of reasons, all of them related to the paths themselves.
By printing the path, we will be able to figure out the actual reason.
Right now we're shooting in the dark.
* deduplicated handling of hrana hello and repreated hello. Code is fully equivalent to the previous form. No change in logic.
* fixed early return
* lazy unwrapping
* made session fields private again
* fmt
* cleaned up nesting in conn
* fmt
allow explain queries without bind parameters
In SQLite, it is invalid to pass a query with a ? or named parameter
without a bind variable, unless you are just trying to explain the
command.
We always fail those queries, but we shouldn't.
This fixes an issue where a db gets restored from bottomless and doesn't
get any writes until shutdown. At this point, the current generation is
the same as the restored one but the graceful shutdown process expects
to wait for that generation to be uploaded which never happens because
there are no writes. This change adds a snapshot generation emit call at
restore time to allow graceful shutdown to happen when there are no
writes without having to checkpoint and upload a new snapshot.
* introduce connection manager
* remove unused wal methods
* remove lock stealer
* Make use of ConnectionManager in LibsqlConnection
it now takes a W: WalWrap instead of a WalManager. This is because we
want to inject the connection manager at the bottom of the wal wrapping
chain.
* add missing deps
* turn ReplicationLogger into a WrapWal
* update spots to to pass wal wrapper instead of wal manager
* remove dbg
* fmt
* fix sqlite3 rust tests
* server: fix conflict on attach update
* add tests and add on conflict to other branch
* only error on server errors
* remove test query
* add more on conflicts fixes
* Revert "add more on conflicts fixes"
This reverts commit 9cef2d320e.
* remove replace for meta store inserts
* wip - passing userauthcontext instead of http headers
* revert comment out
* fixed temp change
* Merge branch 'main' into jw/changing-how-jwt-is-passed-around
* fixed failing tests
* next iteration of unit test fixing
* fixed remaining unit tests
* reverted debug
* removed accidentally added file
* moved str to userauthcontext conversion to from trait
* remove vague comment
* cargo fmt
* cleaned up mod reimportss
* refactored cryptic matching in replica_proxy
* marked potentially duplicate code with // todo dupe #auth
* refactored context to custome errors
* added a factory to produce empty UserAuthContext
* added constructors for UserAuthContext
* switched from try_into to using constructors
* cargo fmt
* added tests for failing cases in parsers
* cargo fmt
* added test for non-asci error
* fixed log message
* removed unnecessary error mapping
* turned context to result
* removing dummy tokens from tests
* cargo fmt + cleanup
---------
Co-authored-by: Julian <julian@Julians-MacBook-Pro.local>
* Hrana: don't close the stream at the end of prepared statement execution
* hrana stream and transaction: check for tokio runtime existence on drop
* standalone::execute_transaction test: raise simulation duration
* hrana: added comments to conditions for stream closing
* hrana: added comments to conditions for stream closing
