Thread: Ghost data from failed FDW transactions?
Hi there!
We have a setup where, for compliance reasons, we hoist a portion of data from several "tenant" databases into a "root" / common / untenanted DB. Through the magic of postgres_fdw, row triggers, and distributed transactions, we automatically hoist the needed columns into the untenanted DB whenever there are changes in any of the tenant DBs. The root DB is never written to outside of these triggers, and act as a sort of cross-tenant index.
I'm scratching my head at a few rows in the root DB, where it seems the corresponding tenant transaction rolled back, but the root DB transaction committed -- there is no row in the tenant but the root DOES have one. I don't have a smoking gun just yet, but this is not the first time we've seen this issue.
Before I jump into particulars, does this sound like expected behavior? We run SERIALIZABLE txn level everywhere (set at the cluster level).
Thanks so much in advance for any insights here!
Here's my setup:
## info + cluster-level flags
GCP cloudsql
postgres version: 15.7default_transaction_isolation: serializable
## The tenant DB:
CREATE TABLE
"devices" (
"orgId" TEXT NOT NULL,
"patientId" TEXT NOT NULL,
"deviceId" TEXT NOT NULL,
"data" JSONB NOT NULL,
PRIMARY KEY ("orgId", "patientId", "deviceId")
);
## The root DB:
CREATE TABLE IF NOT EXISTS "devices" (
"orgId" TEXT NOT NULL,
"patientId" TEXT NOT NULL,
"deviceId" TEXT NOT NULL,
"data" JSONB NOT NULL,
"serialNumber" TEXT NOT NULL GENERATED ALWAYS AS ("data"->>'serialNumber') STORED,
"status" TEXT NOT NULL GENERATED ALWAYS AS ("data"->>'status') STORED,
PRIMARY KEY ("orgId", "patientId", "deviceId"),
CONSTRAINT "deviceIdMatches" CHECK ("data"->>'id' = "deviceId"),
CONSTRAINT "patientIdMatches" CHECK ("data"->>'patientId' = "patientId"),
-- Prevent duplicate serial numbers that are simultaneously "active"
EXCLUDE ( "serialNumber" WITH = ) WHERE ("status" = 'active')
);
## FDW connection from tenant DB to the root:
CREATE EXTENSION IF NOT EXISTS postgres_fdw;
CREATE SERVER IF NOT EXISTS "fdw_server__root" FOREIGN DATA WRAPPER postgres_fdw OPTIONS (
host '${instance-ip-address}',
dbname '${root-db}',
updatable 'true',
truncatable 'false',
keep_connections 'off'
);
CREATE USER MAPPING FOR "${remote-user}" SERVER "fdw_server__root" OPTIONS(
user '${remote-user}', password '$${PGPASSWORD}'
);
CREATE SCHEMA IF NOT EXISTS "rootDb";
GRANT USAGE ON SCHEMA "rootDb" TO "${user-name}";
GRANT SELECT, INSERT, UPDATE, DELETE ON ALL TABLES IN SCHEMA "rootDb" TO "${user-name}";
ALTER DEFAULT PRIVILEGES IN SCHEMA "rootDb" GRANT SELECT, INSERT, UPDATE, DELETE ON TABLES TO "${user-name}";
IMPORT FOREIGN SCHEMA "public" LIMIT TO (devices)
FROM SERVER "fdw_server__root"
INTO "rootDb";
## Trigger setup on the tenant DB, hoisting rows when modified into the root table:
-- Set up a trigger which hoists tenant devices into the rootDb
CREATE OR REPLACE FUNCTION hoist_devices() RETURNS TRIGGER
AS $hoist_devices$
BEGIN
IF (TG_OP IN ('UPDATE', 'DELETE')) THEN
DELETE FROM "rootDb"."devices"
WHERE "orgId" = OLD."orgId"
AND "patientId" = OLD."patientId"
AND "deviceId" = OLD."deviceId";
END IF;
IF (TG_OP IN ('INSERT', 'UPDATE')) THEN
INSERT INTO "rootDb"."devices" ("orgId", "patientId", "deviceId", "data")
SELECT
NEW."orgId",
NEW."patientId",
NEW."deviceId",
NEW."data";
END IF;
RETURN NEW;
END;
$hoist_devices$ LANGUAGE plpgsql;
CREATE OR REPLACE TRIGGER hoist_devices_insert_update_delete
AFTER INSERT OR UPDATE OR DELETE ON "devices"
FOR EACH ROW
EXECUTE FUNCTION hoist_devices();
## A particular endpoint attempts insertions like:
BEGIN;
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")
VALUES (
'org1',
'patient1',
'device1',
'{"id": "device1", "patientId": "patient1", "serialNumber": "12345", "status": "active" }'
);
COMMIT;
--
Jake Biesinger
Jake Biesinger
On Tue, Aug 27, 2024 at 9:03 PM Jacob Biesinger <jake.biesinger@gmail.com> wrote:
I'm scratching my head at a few rows in the root DB, where it seems the corresponding tenant transaction rolled back, but the root DB transaction committed
...
Before I jump into particulars, does this sound like expected behavior?
No, it sounds like something is going wrong. Your setup as described should work to keep both sides in sync.
Through the magic of postgres_fdw, row triggers, and distributed transactions,
Can you expand on "distributed transactions" here?
Cheers,
Greg
On Wed, Aug 28, 2024 at 5:39 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:
On Tue, Aug 27, 2024 at 9:03 PM Jacob Biesinger <jake.biesinger@gmail.com> wrote:I'm scratching my head at a few rows in the root DB, where it seems the corresponding tenant transaction rolled back, but the root DB transaction committed...Before I jump into particulars, does this sound like expected behavior?No, it sounds like something is going wrong. Your setup as described should work to keep both sides in sync.
Well, that's a plus. At least we're (probably) not using it wrong :)
There aren't many details in the docs around failure modes... is there anything there that could cause this issue?
For example, if there were some temporary network outage that prevented communication between the two DBs, would the FDW fail gracefully? Or if there were some disk failure or something after the FDW signals to the remote (root) DB to commit but then the local (tenant) DB failed to commit? We've had a few outages over the years where we hit the `max_connections` setting on the cluster (which would affect both the root + tenant DBs), but connections are held for the duration of both local + remote txns, so doesn't seem like that would affect this. We don't use pgBouncer, either on the client -> DB or as an in-between on the DB -> DB FDW side.
Through the magic of postgres_fdw, row triggers, and distributed transactions,Can you expand on "distributed transactions" here?
I just mean "using the FDW as described". It is magic that we can get proper transactional + ACID semantics in a distributed system. The FDW really helps our use-case.
But to go deeper, we use the javascript knex adapter and some application-level transaction management that automatically retries a transaction N times when it encounters serialization errors. On this particular endpoint, the emitted SQL for the full transaction looks something like:
BEGIN;
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")
VALUES (
'org1',
'patient1',
'device1',
'{"id": "device1", "patientId": "patient1", "serialNumber": "12345", "status": "active" }'
);
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")
VALUES (
'org1',
'patient1',
'device2',
'{"id": "device2", "patientId": "patient1", "serialNumber": "67890", "status": "active" }'
);
SELECT * FROM "rootDb"."assets";
-- execute some logic client-side, nothing touching the DB
UPDATE "rootDb"."assets" WHERE ...;
COMMIT;
With I guess the maybe-relevant bits here being that we do some additional reading + writing to the remote / root DB (other tables) subsequently as part of the same txn. The JS driving this also has the unfortunate shortcoming that the two `INSERT` statements run "in parallel", meaning there's a race to execute them (serially) through their shared txn/connection. The ordering shouldn't matter, but this also means that error handling (e.g., when there is a conflict with the `EXCLUDE "serialNumber"` constraint) may not stop the second `INSERT` statement from being attempted (and rejected by postgres). But I think that's all client-side details that shouldn't affect the FDW txn semantics, right?
--
Jake Biesinger
On Wed, Aug 28, 2024 at 5:39 AM Greg Sabino Mullane <htamfids@gmail.com> wrote:
On Tue, Aug 27, 2024 at 9:03 PM Jacob Biesinger <jake.biesinger@gmail.com> wrote:I'm scratching my head at a few rows in the root DB, where it seems the corresponding tenant transaction rolled back, but the root DB transaction committed...Before I jump into particulars, does this sound like expected behavior?No, it sounds like something is going wrong. Your setup as described should work to keep both sides in sync.Through the magic of postgres_fdw, row triggers, and distributed transactions,Can you expand on "distributed transactions" here?Cheers,Greg
On Aug 28, 2024, at 10:18 AM, Jacob Biesinger <jake.biesinger@gmail.com> wrote:But to go deeper, we use the javascript knex adapter and some application-level transaction management that automatically retries a transaction N times when it encounters serialization errors. On this particular endpoint, the emitted SQL for the full transaction looks something like:BEGIN;INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")VALUES ('org1','patient1','device1','{"id": "device1", "patientId": "patient1", "serialNumber": "12345", "status": "active" }');INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")VALUES ('org1','patient1','device2','{"id": "device2", "patientId": "patient1", "serialNumber": "67890", "status": "active" }');SELECT * FROM "rootDb"."assets";-- execute some logic client-side, nothing touching the DBUPDATE "rootDb"."assets" WHERE ...;COMMIT;
BEGIN;
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data")
VALUES (
'org1',
'patient1',
'device1',
'{"id": "device1", "patientId": "patient1", "serialNumber": "12345", "status": "active" }’),
(
'org1',
'patient1',
'device2',
'{"id": "device2", "patientId": "patient1", "serialNumber": "67890", "status": "active" }'
)
Any value in supplying a single insert statement a la (less back and forth perhaps?):
Yes, absolutely that would be better. This particular endpoint has some ancient + crufty code backing it (migrated from a NoSQL DB with a db-agnostic shim that we're slowly replacing). The old code likes doing things client-side instead of being sane about SQL semantics.
But I don't think that would affect the issue here, right?
Jake Biesinger
Any updates on this?
A few replies from me inline:
On Wed, Aug 28, 2024 at 12:18 PM Jacob Biesinger <jake.biesinger@gmail.com> wrote:
There aren't many details in the docs around failure modes... is there anything there that could cause this issue?
Nothing that I know of, but it's possible there is some sort of weird race condition/bug. But frankly, it's far more likely to be some sort of application bug / missed behavior. If you could trigger it at will by developing a self-contained test, that would be ideal.
The ordering shouldn't matter, but this also means that error handling (e.g., when there is a conflict with the `EXCLUDE "serialNumber"` constraint) may not stop the second `INSERT` statement from being attempted (and rejected by postgres). But I think that's all client-side details that shouldn't affect the FDW txn semantics, right?
Correct, but it's hard to say for sure without seeing the code or knowing exactly how the serialization errors are being handled. I'm not clear on how the inserts are run "in parallel" if your flow is:
BEGIN;
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data") ...
INSERT INTO "devices" ("orgId", "patientId", "deviceId", "data") ...
SELECT * FROM "rootDb"."assets";
-- execute some logic client-side, nothing touching the DB
UPDATE "rootDb"."assets" WHERE ...;
COMMIT;
Perhaps expand on that, because it almost sounds like you have two processes trying to talk to the same backend at once?
Cheers,
Greg