EDB Postgres Distributed is an active/active or multi-master DBMS. If
used asynchronously, writes to the same or related rows from multiple
different nodes can result in data conflicts when using standard data
types.
Conflicts aren’t errors. In most cases, they are events that PGD can
detect and resolve as they occur. Resolving them depends on the nature
of the application and the meaning of the data, so it’s important for
PGD to provide the application a range of choices for how to resolve
conflicts.
By default, conflicts are resolved at the row level. When changes from
two nodes conflict, PGD picks either the local or remote tuple and the
discards the other. For example, the commit timestamps might be compared
for the two conflicting changes and the newer one kept. This approach
ensures that all nodes converge to the same result and establishes
commit-order-like semantics on the whole cluster.
Column-level conflict detection and resolution is available with PGD,
described in CLCD .
Some clustering systems use distributed lock mechanisms to prevent
concurrent access to data. These can perform reasonably when servers are
very close to each other but can’t support geographically distributed
applications where very low latency is critical for acceptable
performance.
Distributed locking is essentially a pessimistic approach. PGD advocates
an optimistic approach, which is to avoid conflicts where possible but
allow some types of conflicts to occur and resolve them when they arise.
Types of conflict
PRIMARY KEY or UNIQUE conflicts
The most common conflicts are row conflicts, where two operations affect
a row with the same key in ways they can’t on a single node. PGD can
detect most of those and applies the update_if_newer conflict
resolver.
Row conflicts include:
INSERT versus INSERT
UPDATE versus UPDATE
UPDATE versus DELETE
INSERT versus UPDATE
INSERT versus DELETE
DELETE versus DELETE
The view bdr.node_conflict_resolvers provides information on how
conflict resolution is currently configured for all known conflict
types.
INSERT/INSERT conflicts
The most common conflict, INSERT /INSERT , arises where
INSERT operations on two different nodes create a tuple with the
same PRIMARY KEY values (or if no PRIMARY KEY exists, the same
values for a single UNIQUE constraint).
PGD handles this situation by retaining the most recently inserted tuple
of the two according to the originating node’s timestamps. (A
user-defined conflict handler can override this behavior.)
This conflict generates the insert_exists conflict type, which is by
default resolved by choosing the newer row, based on commit time, and
keeping only that one (update_if_newer resolver). You can configure
other resolvers. See Enabling and disabling column-level conflict resolution for details.
To resolve this conflict type, you can also use column-level conflict
resolution and user-defined conflict triggers.
- You can effectively eliminate this type of conflict by using
PGD global sequences .
INSERT operations that violate multiple UNIQUE constraints
An INSERT /INSERT conflict can violate more than one UNIQUE
constraint, of which one might be the PRIMARY KEY . If a new row
violates more than one UNIQUE constraint and that results in a
conflict against more than one other row, then applying the replication
change produces a multiple_unique_conflicts conflict.
In case of such a conflict, you must remove some rows for replication to
continue. Depending on the resolver setting for
multiple_unique_conflicts , the apply process either exits with
error, skips the incoming row, or deletes some of the rows. The deletion
tries to preserve the row with the correct PRIMARY KEY and delete
the others.
Warning
In case of multiple rows conflicting this way, if the result of conflict resolution is to proceed with the insert operation, some of the data is always deleted.
You can also define a different behavior using a conflict trigger.
UPDATE/UPDATE conflicts
Where two concurrent UPDATE operations on different nodes change the
same tuple but not its PRIMARY KEY , an UPDATE /UPDATE
conflict can occur on replay.
These can generate different conflict kinds based on the configuration
and situation. If the table is configured with Row version conflict detection , then the
original (key) row is compared with the local row. If they’re different,
the update_differing conflict is generated. When using Origin conflict detection
, the origin of the row is checked. (The origin is the node that the
current local row came from.) If that changed, the
update_origin_change conflict is generated. In all other cases, the
UPDATE is normally applied without generating a conflict.
Both of these conflicts are resolved the same way as insert_exists ,
described in INSERT/INSERT conflicts .
UPDATE conflicts on the PRIMARY KEY
PGD can’t currently perform conflict resolution where the
PRIMARY KEY is changed by an UPDATE operation. You can update
the primary key, but you must ensure that no conflict with existing
values is possible.
Conflicts on the update of the primary key are Divergent conflicts and
require manual intervention.
Updating a primary key is possible in Postgres, but there are issues in
both Postgres and PGD.
A simple schema provides an example that explains:
CREATE TABLE pktest (pk integer primary key, val integer);
INSERT INTO pktest VALUES (1,1);
Updating the Primary Key column is possible, so this SQL succeeds:
UPDATE pktest SET pk=2 WHERE pk=1;
However, suppose the table has multiple rows:
INSERT INTO pktest VALUES (3,3);
Some UPDATE operations succeed:
UPDATE pktest SET pk=4 WHERE pk=3;
SELECT * FROM pktest;
pk | val
- ---+-----
2 | 1
4 | 3
(2 rows)
Other UPDATE operations fail with constraint errors:
UPDATE pktest SET pk=4 WHERE pk=2;
ERROR: duplicate key value violates unique constraint "pktest_pkey"
DETAIL: Key (pk)=(4) already exists
So for Postgres applications that update primary keys, be careful to
avoid runtime errors, even without PGD.
With PGD, the situation becomes more complex if UPDATE operations are
allowed from multiple locations at same time.
Executing these two changes concurrently works:
node1: UPDATE pktest SET pk=pk+1 WHERE pk = 2;
node2: UPDATE pktest SET pk=pk+1 WHERE pk = 4;
SELECT * FROM pktest;
pk | val
- ---+-----
3 | 1
5 | 3
(2 rows)
Executing these next two changes concurrently causes a divergent error,
since both changes are accepted. But applying the changes on the other
node results in update_missing conflicts.
node1: UPDATE pktest SET pk=1 WHERE pk = 3;
node2: UPDATE pktest SET pk=2 WHERE pk = 3;
This scenario leaves the data different on each node:
node1:
SELECT * FROM pktest;
pk | val
- ---+-----
1 | 1
5 | 3
(2 rows)
node2:
SELECT * FROM pktest;
pk | val
- ---+-----
2 | 1
5 | 3
(2 rows)
You can identify and resolve this situation using Data verification with LiveCompare .
Concurrent conflicts present problems. Executing these two changes
concurrently isn’t easy to resolve:
node1: UPDATE pktest SET pk=6, val=8 WHERE pk = 5;
node2: UPDATE pktest SET pk=6, val=9 WHERE pk = 5;
Both changes are applied locally, causing a divergence between the
nodes. But the apply on the target fails on both nodes with a duplicate
key-value violation error. This error causes the replication to halt and
requires manual resolution.
You can avoid this duplicate key violation error, and replication
doesn’t break, if you set the conflict_type update_pkey_exists to
skip , update , or update_if_newer . This approach can still
lead to divergence depending on the nature of the update.
You can avoid divergence in cases where the same old key is being
updated by the same new key concurrently by setting
update_pkey_exists to update_if_newer . However, in certain
situations, divergence occurs even with update_if_newer , namely
when two different rows both are updated concurrently to the same new
primary key.
As a result, we recommend strongly against allowing primary key UPDATE
operations in your applications, especially with PGD. If parts of your
application change primary keys, then to avoid concurrent changes, make
those changes using Eager Replication.
Warning
In case the conflict resolution of update_pkey_exists conflict results in update, one of the rows is always deleted.
UPDATE operations that violate multiple UNIQUE constraints
Like INSERT operations that violate multiple UNIQUE constraints , when an incoming UPDATE violates more than one
UNIQUE index (or the PRIMARY KEY ), PGD raises a
multiple_unique_conflicts conflict.
PGD supports deferred unique constraints. If a transaction can commit on
the source, then it applies cleanly on target, unless it sees conflicts.
However, you can’t use a deferred primary key as a REPLICA IDENTITY, so
the use cases are already limited by that and the warning about using
multiple unique constraints.
UPDATE/DELETE conflicts
One node can update a row that another node deletes at ths same time. In
this case an UPDATE /DELETE conflict can occur on replay.
If the deleted row is still detectable (the deleted row wasn’t removed
by VACUUM ), the update_recently_deleted conflict is generated.
By default, the UPDATE is skipped, but you can configure the
resolution for this. See Enabling and disabling column-level conflict resolution for details.
The database can clean up the deleted row by the time the UPDATE is
received in case the local node is lagging behind in replication. In
this case, PGD can’t differentiate between UPDATE /DELETE
conflicts and INSERT/UPDATE conflicts . It generates the update_missing
conflict.
Another type of conflicting DELETE and UPDATE is a DELETE
that comes after the row was updated locally. In this situation, the
outcome depends on the type of conflict detection used. When using the
default, Origin conflict detection , no conflict is detected, leading to the
DELETE being applied and the row removed. If you enable
The default resolution for this conflict type is to apply the DELETE
and remove the row. However, you can configure this or a conflict
trigger can handled it.
INSERT/UPDATE conflicts
When using the default asynchronous mode of operation, a node might
receive an UPDATE of a row before the original INSERT was
received. This can happen only when three or more nodes are active (see
When this happens, the update_missing conflict is generated. The
default conflict resolver is insert_or_skip , though you can use
insert_or_error or skip instead. Resolvers that do
insert-or-action first try to INSERT a new row based on data from
the UPDATE when possible (when the whole row was received). For
reconstructing the row to be possible, the table either needs to have
REPLICA IDENTITY FULL or the row must not contain any toasted data.
See TOAST support details for more info about toasted data.
INSERT/DELETE conflicts
Similar to the INSERT /UPDATE conflict, the node might also
receive a DELETE operation on a row for which it didn’t yet receive
an INSERT . This is again possible only with three or more nodes set
up (see Conflicts with three or more nodes ).
PGD can’t currently detect this conflict type. The INSERT operation
doesn’t generate any conflict type, and the INSERT is applied.
The DELETE operation always generates a delete_missing conflict,
which is by default resolved by skipping the operation.
DELETE/DELETE conflicts
A DELETE /DELETE conflict arises when two different nodes
concurrently delete the same tuple.
This scenario always generates a delete_missing conflict, which is
by default resolved by skipping the operation.
This conflict is harmless since both DELETE operations have the same
effect. You can safely ignroe one of them.
Conflicts with three or more nodes
If one node inserts a row that’s then replayed to a second node and
updated there, a third node can receive the UPDATE from the second
node before it receives the INSERT from the first node. This
scenario is an INSERT /UPDATE conflict.
These conflicts are handled by discarding the UPDATE , which can
lead to different data on different nodes. These are Divergent conflicts .
This conflict type can happen only with three or more masters. At least
two masters must be actively writing.
Also, the replication lag from node 1 to node 3 must be high enough to
allow the following sequence of actions:
node 2 receives INSERT from node 1
node 2 performs UPDATE
node 3 receives UPDATE from node 2
node 3 receives INSERT from node 1
Using insert_or_error (or in some cases the insert_or_skip
conflict resolver for the update_missing conflict type) is a viable
mitigation strategy for these conflicts. However, enabling this option
opens the door for INSERT /DELETE conflicts:
node 1 performs UPDATE
node 2 performs DELETE
node 3 receives DELETE from node 2
node 3 receives UPDATE from node 1, turning it into an INSERT
If these are problems, we recommend tuning freezing settings for a table
or database so that they’re correctly detected as
update_recently_deleted .
Another alternative is to use Effects of Eager Replication in general to prevent these conflicts.
INSERT /DELETE conflicts can also occur with three or more
nodes. Such a conflict is identical to INSERT /UPDATE except
with the UPDATE replaced by a DELETE . This can result in a
delete_missing conflict.
PGD could choose to make each INSERT into a check-for-recently
deleted, as occurs with an update_missing conflict. However, the
cost of doing this penalizes the majority of users, so at this time it
instead logs delete_missing .
Future releases will automatically resolve INSERT /DELETE
anomalies by way of rechecks using Data verification with LiveCompare when
delete_missing conflicts occur. Applications can perform these
manually by checking the bdr.conflict_history_summary view.
These conflicts can occur in two main problem use cases:
INSERT followed rapidly by a DELETE , as can be used in
queuing applications
Any case where the primary key identifier of a table is reused
Neither of these cases is common. We recommend not replicating the
affected tables if these problem use cases occur.
PGD has problems with the latter case because PGD relies on the
uniqueness of identifiers to make replication work correctly.
Applications that insert, delete, and then later reuse the same unique
identifiers can cause difficulties. This is known as the ABA problem
. PGD has no way of knowing whether the rows are the current row, the
last row, or much older rows.
Unique identifier reuse is also a business problem, since it is prevents
unique identification over time, which prevents auditing, traceability,
and sensible data quality. Applications don’t need to reuse unique
identifiers.
Any identifier reuse that occurs in the time interval it takes for
changes to pass across the system causes difficulties. Although that
time might be short in normal operation, down nodes can extend that
interval to hours or days.
We recommend that applications don’t reuse unique identifiers. If they
do, take steps to avoid reuse in less than a year.
This problem doesn’t occur in applications that use sequences or UUIDs.
Foreign key constraint conflicts
Conflicts between a remote transaction being applied and existing local
data can also occur for FOREIGN KEY (FK) constraints.
PGD applies changes with session_replication_role = 'replica' , so
foreign keys aren’t rechecked when applying changes. In an active/active
environment, this situation can result in FK violations if deletes occur
to the referenced table at the same time as inserts into the referencing
table. This scenario is similar to an INSERT /DELETE conflict.
In single-master Postgres, any INSERT /UPDATE that refers to a
value in the referenced table must wait for DELETE operations to
finish before they can gain a row-level lock. If a DELETE removes a
referenced value, then the INSERT /UPDATE fails the FK check.
In multi-master PGD. there are no inter-node row-level locks. An
INSERT on the referencing table doesn’t wait behind a DELETE on
the referenced table, so both actions can occur concurrently. Thus an
INSERT /UPDATE on one node on the referencing table can use a
value at the same time as a DELETE on the referenced table on
another node. The result, then, is a value in the referencing table
that’s no longer present in the referenced table.
In practice, this situation occurs if the DELETE operations occurs
on referenced tables in separate transactions from DELETE operations
on referencing tables, which isn’t a common operation.
In a parent-child relationship such as Orders -> OrderItems, it isn’t
typical to do this. It’s more likely to mark an OrderItem as canceled
than to remove it completely. For reference/lookup data, it’s unusual to
completely remove entries at the same time as using those same values
for new fact data.
While dangling FKs are possible, the risk of this in general is very
low. Thus PGD doesn’t impose a generic solution to cover this case. Once
you understand the situation in which this occurs, two solutions are
possible.
The first solution is to restrict the use of FKs to closely related
entities that are generally modified from only one node at a time, are
infrequently modified, or where the modification’s concurrency is
application mediated. This approach avoids any FK violations at the
application level.
The second solution is to add triggers to protect against this case
using the PGD-provided functions bdr.ri_fkey_trigger() and
bdr.ri_fkey_on_del_trigger() . When called as BEFORE triggers,
these functions use FOREIGN KEY information to avoid FK anomalies by
setting referencing columns to NULL, much as if you had a SET NULL
constraint. This approach rechecks all FKs in one trigger, so you need
to add only one trigger per table to prevent FK violation.
As an example, suppose you have two tables: Fact and RefData. Fact has
an FK that references RefData. Fact is the referencing table, and
RefData is the referenced table. You need to add one trigger to each
table.
Add a trigger that sets columns to NULL in Fact if the referenced row in
RefData was already deleted:
CREATE TRIGGER bdr_replica_fk_iu_trg
BEFORE INSERT OR UPDATE ON fact
FOR EACH ROW
EXECUTE PROCEDURE bdr.ri_fkey_trigger();
ALTER TABLE fact
ENABLE REPLICA TRIGGER bdr_replica_fk_iu_trg;
Add a trigger that sets columns to NULL in Fact at the time a DELETE
occurs on the RefData table:
CREATE TRIGGER bdr_replica_fk_d_trg
BEFORE DELETE ON refdata
FOR EACH ROW
EXECUTE PROCEDURE bdr.ri_fkey_on_del_trigger();
ALTER TABLE refdata
ENABLE REPLICA TRIGGER bdr_replica_fk_d_trg;
Adding both triggers avoids dangling foreign keys.
TRUNCATE conflicts
TRUNCATE behaves similarly to a DELETE of all rows but performs
this action by physically removing the table data rather than row-by-row
deletion. As a result, row-level conflict handling isn’t available, so
TRUNCATE commands don’t generate conflicts with other DML actions,
even when there’s a clear conflict.
As a result, the ordering of replay can cause divergent changes if
another DML is executed concurrently on other nodes to the TRUNCATE
.
You can take one of the following actions:
Ensure TRUNCATE isn’t executed alongside other concurrent DML.
Rely on Data verification with LiveCompare to highlight any such inconsistency.
Replace TRUNCATE with a DELETE statement with no WHERE
clause. This approach is likely to have poor performance on larger
tables.
Set bdr.truncate_locking = 'on' to set the TRUNCATE command’s
locking behavior. This setting determines whether TRUNCATE obeys
the bdr.ddl_locking setting. This isn’t the default behavior for
TRUNCATE since it requires all nodes to be up. This configuration
might not be possible or wanted in all cases.
Exclusion constraint conflicts
PGD doesn’t support exclusion constraints and prevents their creation.
If an existing standalone database is converted to a PGD database, then
drop all exclusion constraints manually.
In a distributed asynchronous system, you can’t ensure that no set of
rows that violate the constraint exists because all transactions on
different nodes are fully isolated. Exclusion constraints lead to replay
deadlocks where replay can’t progress from any node to any other node
because of exclusion constraint violations.
If you force PGD to create an exclusion constraint, or you don’t drop
existing ones when converting a standalone database to PGD, expect
replication to break. To get it to progress again, remove or alter the
local tuples that an incoming remote tuple conflicts with so that the
remote transaction can be applied.
Data conflicts for roles and tablespace differences
Conflicts can also arise where nodes have global (Postgres-system-wide)
data, like roles, that differ. This conflict can result in
operations—mainly DDL —that can run successfully and commit on one
node but then fail to apply to other nodes.
For example, node1 might have a user named fred, and that user wasn’t
created on node2. If fred on node1 creates a table, the table is
replicated with its owner set to fred. When the DDL command is applied
to node2, the DDL fails because there’s no user named fred. This failure
generates an error in the Postgres logs.
Administrator intervention is required to resolve this conflict by
creating the user fred in the database where PGD is running. You can set
bdr.role_replication = on to resolve this in future.
Lock conflicts and deadlock aborts
Because PGD writer processes operate much like normal user sessions,
they’re subject to the usual rules around row and table locking. This
can sometimes lead to PGD writer processes waiting on locks held by user
transactions or even by each other.
Relevant locking includes:
Explicit table-level locking (LOCK TABLE ... ) by user sessions
Explicit row-level locking (SELECT ... FOR UPDATE/FOR SHARE ) by
user sessions
Implicit locking because of row UPDATE , INSERT , or
DELETE operations, either from local activity or from replication
from other nodes
A PGD writer process can deadlock with a user transaction, where the
user transaction is waiting on a lock held by the writer process and
vice versa. Two writer processes can also deadlock with each other.
Postgres’s deadlock detector steps in and terminates one of the problem
transactions. If the PGD writer process is terminated, it retries and
generally succeeds.
All these issues are transient and generally require no administrator
action. If a writer process is stuck for a long time behind a lock on an
idle user session, the administrator can terminate the user session to
get replication flowing again. However, this is no different from a user
holding a long lock that impacts another user session.
Use of the log_lock_waits facility in Postgres can help identify locking
related replay stalls.
Divergent conflicts
Divergent conflicts arise when data that should be the same on different
nodes differs unexpectedly. Divergent conflicts shouldn’t occur, but not
all such conflicts can be reliably prevented at the time of writing.
Changing the PRIMARY KEY of a row can lead to a divergent conflict
if another node changes the key of the same row before all nodes have
replayed the change. Avoid changing primary keys, or change them only on
one designated node.
Divergent conflicts involving row data generally require administrator
action to manually adjust the data on one of the nodes to be consistent
with the other one. Such conflicts don’t arise so long as you use PGD as
documented and avoid settings or functions marked as unsafe.
The administrator must manually resolve such conflicts. You might need
to use the advanced options such as bdr.ddl_replication and
bdr.ddl_locking depending on the nature of the conflict. However,
careless use of these options can make things much worse and create a
conflict that generic instructions can’t address.
TOAST support details
- Postgres uses out-of-line storage for larger columns called
TOAST .
The TOAST values handling in logical decoding (which PGD is built on top
of) and logical replication is different from inline data stored as part
of the main row in the table.
The TOAST value is logged into the transaction log (WAL) only if the
value changed. This can cause problems, especially when handling UPDATE
conflicts, because an UPDATE statement that didn’t change a value of
a toasted column produces a row without that column. As mentioned in
is resolved using insert_or_error and there are missing TOAST
columns.
However, more subtle issues than this one occur in case of concurrent
workloads with asynchronous replication. (Eager transactions aren’t
affected.) Imagine, for example, the following workload on an EDB
Postgres Distributed cluster with three nodes called A, B, and C:
On node A: txn A1 does an UPDATE SET col1 = ‘toast data…’ and commits
first.
On node B: txn B1 does UPDATE SET other_column = ‘anything else’; and
commits after A1.
On node C: the connection to node A lags behind.
On node C: txn B1 is applied first, it misses the TOASTed column in
col1, but gets applied without conflict.
On node C: txn A1 conflicts (on update_origin_change) and is skipped.
Node C misses the toasted data from A1 forever.
This scenario isn’t usually a problem when using PGD. (It is when using
either built-in logical replication or plain pglogical for
multi-master.) PGD adds its own logging of TOAST columns when it detects
a local UPDATE to a row that recently replicated a TOAST column
modification and the local UPDATE isn’t modifying the TOAST. Thus
PGD prevents any inconsistency for toasted data across different nodes.
This situation causes increased WAL logging when updates occur on
multiple nodes, that is, when origin changes for a tuple. Additional WAL
overhead is zero if all updates are made from a single node, as is
normally the case with PGD AlwaysOn architecture.
Note
Running VACUUM FULL or CLUSTER on just the TOAST table without doing same on the main table removes metadata needed for the extra logging to work. This means that, for a short period after such a statement, the protection against these concurrency issues isn’t present.
Warning
The additional WAL logging of TOAST is done using the BEFORE UPDATE trigger on standard Postgres. This trigger must be sorted alphabetically last based on trigger name among all BEFORE UPDATE triggers on the table. It’s prefixed with zzzz_bdr_ to make this easier, but make sure you don’t create any trigger with a name that sorts after it. Otherwise you won’t have the protection against the concurrency issues.
For the insert_or_error conflict resolution, the use of
REPLICA IDENTITY FULL is still required.
None of these problems associated with toasted columns affect tables
with REPLICA IDENTITY FULL . This setting always logs a toasted
value as part of the key since the whole row is considered to be part of
the key. PGD can reconstruct the new row, filling the missing data from
the key row. As a result, using REPLICA IDENTITY FULL can increase
WAL size significantly.
