do not use prepared statements for mutations

CHANGELOG.md

@@ -8,6 +8,7 @@
 
 - server: add new `--conn-lifetime` and `HASURA_GRAPHQL_PG_CONN_LIFETIME` options for expiring connections after some amount of active time (#5087)
 - server: shrink libpq connection request/response buffers back to 1MB if they grow beyond 2MB, fixing leak-like behavior on active servers (#5087)
+- server: disable prepared statements for mutations as we end up with single-use objects which result in excessive memory consumption for mutation heavy workloads (#5255)
 - console: allow configuring statement timeout on console RawSQL page (close #4998) (#5045)
 - docs: add note for managed databases in postgres requirements (close #1677, #3783) (#5228)
 - docs: add 1-click deployment to Nhost page to the deployment guides (#5180)
@@ -119,7 +120,7 @@ Read more about the session argument for computed fields in the [docs](https://h
 A new `seeds` command is introduced in CLI, this will allow managing seed migrations as SQL files
 
 #### Creating seed
-```                                                        
+```
 # create a new seed file and use editor to add SQL content
 hasura seed create new_table_seed
 

server/src-lib/Hasura/RQL/DML/Mutation.hs

@@ -72,6 +72,20 @@ mutateAndReturn (Mutation qt (cte, p) mutationOutput allCols remoteJoins strfyNu
     sqlQuery = Q.fromBuilder $ toSQL $
                mkMutationOutputExp qt allCols Nothing cte mutationOutput strfyNum
 
+{- Note: [Prepared statements in Mutations]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+The SQL statements we generate for mutations seem to include the actual values
+in the statements in some cases which pretty much makes them unfit for reuse
+(Handling relationships in the returning clause is the source of this
+complexity). Further, `PGConn` has an internal cache which maps a statement to
+a 'prepared statement id' on Postgres. As we prepare more and more single-use
+SQL statements we end up leaking memory both on graphql-engine and Postgres
+till the connection is closed. So a simpler but very crude fix is to not use
+prepared statements for mutations. The performance of insert mutations
+shouldn't be affected but updates and delete mutations with complex boolean
+conditions **might** see some degradation.
+-}
+
 mutateAndSel
   :: (HasVersion, MonadTx m, MonadIO m)
   => Mutation -> m EncJSON
@@ -92,7 +106,8 @@ executeMutationOutputQuery
 executeMutationOutputQuery query prepArgs = \case
   Nothing ->
     runIdentity . Q.getRow
-      <$> liftTx (Q.rawQE dmlTxErrorHandler query prepArgs True)
+      -- See Note [Prepared statements in Mutations]
+      <$> liftTx (Q.rawQE dmlTxErrorHandler query prepArgs False)
   Just (remoteJoins, (httpManager, reqHeaders, userInfo)) ->
     executeQueryWithRemoteJoins httpManager reqHeaders userInfo query prepArgs remoteJoins
 
@@ -105,7 +120,8 @@ mutateAndFetchCols
   -> Q.TxE QErr (MutateResp TxtEncodedPGVal)
 mutateAndFetchCols qt cols (cte, p) strfyNum =
   Q.getAltJ . runIdentity . Q.getRow
-    <$> Q.rawQE dmlTxErrorHandler (Q.fromBuilder sql) (toList p) True
+    -- See Note [Prepared statements in Mutations]
+    <$> Q.rawQE dmlTxErrorHandler (Q.fromBuilder sql) (toList p) False
   where
     aliasIden = Iden $ qualObjectToText qt <> "__mutation_result"
     tabFrom = FromIden aliasIden