PostgreSQL Cluster Switch-Over streaming issues!

How to handle replication after a failover of PostgreSQL Cluster from Primary to Secondary and back to Primary for the data streaming using Debezium and Kafka?

In a standard PostgreSQL and Debezium architecture, Debezium will not automatically pick up the new WAL and resume seamlessly after a true failover and failback. It almost always requires manual intervention or a reset.

The reason for this comes down to how PostgreSQL handles logical replication.

Ø  The Core Problem: Logical Replication Slots

Debezium does not read the raw physical WAL files directly. It relies on a Logical Replication Slot (usually using the pgoutput plugin) created on the source database.

1.     Slots are Node-Specific: By default, logical replication slots are tied exclusively to the primary server. They are not physically replicated to the DR standby node alongside the data.

2.     The Failover (To DR): When you failed over to the DR node, that DR database did not have Debezium's logical replication slot. The Debezium connector immediately threw an error and failed.

3.     The Timeline Change: When a DR node is promoted to Primary, PostgreSQL increments its "timeline."

4.     The Failback (To Primary): When you failed back to the original primary, the primary had to reconcile with the DR node's new timeline. In most failback scenarios (especially if pg_rewind or a fresh backup was used), the original logical replication slot is destroyed or rendered invalid because the Log Sequence Numbers (LSNs) no longer perfectly align.

Ø  How to Resolve and Reset

If the Debezium connector is currently in a FAILED state, then we will likely need to reset it. Here is the standard recovery path:

ü          Step 1: Check the Slot on the Primary First, verify if the slot survived the failback on the source PostgreSQL database:

Query:

SELECT slot_name, plugin, slot_type, active, restart_lsn

FROM pg_replication_slots

WHERE slot_name = 'your_debezium_slot_name';

·       If the query returns zero rows, the slot was destroyed.

·       If it returns a row but active is false and Debezium refuses to connect, the LSNs are out of sync.

ü          Step 2: Drop the Corrupted Slot (If it exists) If the slot is there but broken, we must drop it manually so Debezium can recreate it:

Query:

               SELECT pg_drop_replication_slot('your_debezium_slot_name');

ü          Step 3: Reset the Debezium Connector Because the LSN pointers have changed and the slot is gone, we cannot just restart the connector. 

            We have to force Debezium to take a new initial snapshot to ensure zero data loss.

·       Delete the connector's offsets in the Kafka Connect offsets topic (or via the Kafka Connect REST API if you are using a tool that supports offset deletion).

·       Reconfigure the Debezium connector with snapshot.mode set to initial (or always).

·       Restart the connector. It will recreate the replication slot, perform a fresh SELECT * baseline read of your tables, and then begin streaming the new WAL changes.

 

 

PostgreSQL Databases Historical Data Archival and Storage!

Hot, Warm, Cold: Architecting a 3-Tier Data Lifecycle for PostgreSQL

Managing data of multi tera bytes requires a planned separation of live data from the warm, cold storage. Keeping terabytes of historical data on premium database disks, even if the partitions are detached and no longer impacting query planner overhead besides impacts expensive IOPS and inflates our backup/restore RTO windows.

Implementing a robust historical data archival strategy is one of the most impactful architectural improvements you can make for a large-scale PostgreSQL estate. It directly solves the pain points of bloated indexes, extended backup windows, and the aggressive locking issues, we experience in our day-to-day work loads.

I'm considering a usecase of 13 servers with ~1TB of raw data and ~750GB of archival data per each database/server on average, the most robust architecture shifts this data off the relational engine and into highly compressed, immutable Object Storage in Cloud storage like AWS S3 or Azure BLOB Storage using a middle tier for OLAP like ClickHouse.

Below mentioned options provides a comprehensive blueprint for the extraction, storage, and retrieval phases, along with the estimated storage footprint and cost projections.

Phase 1: The Extraction & Transformation Strategy

Once a partition rolls past the historical retention period and is detached from the main table, it must be extracted efficiently without causing I/O spikes that compete with our active DML operations.

• The Format (Parquet vs. CSV/ZSTD): * Zstandard (ZSTD) Compressed CSV: If this data is strictly for compliance and rarely touched, streaming the detached partition via COPY to a ZSTD-compressed CSV is the fastest and most CPU-efficient method.

o   Apache Parquet: If there is a chance the business or analytics teams will need to query this historical data, extract it into Parquet format. Parquet is columnar, highly compressed, and can be queried directly in object storage.

·       The Workflow:

1. Detach the partition from the parent table (instantly removes it from live query plans).
2. Execute a background worker script to export the detached table to the target format into AWS S3 or Azure blob storage.
3. Stream the output directly to Object Storage (e.g., Azure Blob Storage or AWS S3) to bypass local disk staging.
4. Verify the checksum of the uploaded data before dropping from Source.
5. DROP TABLE on the PostgreSQL server to reclaim the premium block storage of ~500GB on each DB Server. 

Phase 2: Storage & Compression Estimates

Relational database data is highly compressible. When you extract 10TB of raw PostgreSQL data (excluding the B-Tree indexes), the actual storage footprint drops dramatically.

Assuming a standard compression ratio of 4:1 (75% reduction) using ZSTD or Parquet:

Metric/Parameter	Size Per Server	Total (13 Servers)
Raw PostgreSQL Data	~750 GB	~10 TB
Index Overhead Removed	~100 GB	2.5 TB
Data Payload to Compress	~500 GB	7.5 TB
Estimated Target Storage (Compressed)	~150 GB	~2 TB

 

Phase 3: The Cost Estimates & Tiering

By shifting this data off Premium SSDs (which typically cost around $150 to $200 per provisioned TB per month depending on the cloud provider or SAN tier) and moving it to Object Storage, the ROI is immediate.

Here is the estimated monthly cost to store the resulting ~2 TB of compressed archive data across standard cloud object storage tiers:

Storage Tier	Best For	Approx. Monthly Cost (2 TB)	Retrieval (per-GB) Characteristics
Hot / Standard Object	Data queried frequently	~$50.00 / month	Millisecond access. Standard retrieval costs.
Cool / Infrequent Access	Queried once a month	~$20.00 / month	Millisecond access. Higher retrieval fee.
Cold / Archive / Glacier	Audits only (1-2x a year)	~$2.50 / month	Hours to retrieve. Highest retrieval fees.

 

Phase 4: The Retrieval Architecture

When compliance or business users inevitably ask for data from this Historical Archival period, We need to provide the same without restoring 1TB back into our live transactional database.

• Option A (The Native FDW): If the data was archived as CSV, you can temporarily create a file_fdw (Foreign Data Wrapper) or use s3_fdw/azure_storage_fdw to map the remote file as a read-only table in a staging PostgreSQL instance.
• Option B (The Data Lake Approach): If the data was archived as Parquet, point an analytical query engine like DuckDB or a ClickHouse DB directly at the storage bucket. This allows you to run standard SQL against the deep archive without touching our primary PostgreSQL clusters.

OLAP phase: 

To bridge the gap between active transactional data and deep archival storage, we have implemented ClickHouse as a dedicated OLAP tier. This ensures high-performance, low-latency querying for frequently accessed historical data, mitigating the latency, indexing, and functional limitations of querying cold data directly from S3 object storage.

This can be best achieved through an event-driven Change Data Capture (CDC) pipeline that transforms raw transactional logs into query-ready analytical states.

Here are the critical architectural steps to build this pipeline:

• Source Configuration (PostgreSQL): Configure the database for logical decoding by setting wal_level = logical, establishing a publication for the specific tables, and creating a logical replication slot to track WAL consumption.
• Event Capture (Debezium): Attach a CDC connector to the replication slot to read the WAL stream, translating every INSERT, UPDATE, and DELETE into a structured event payload (typically JSON or Avro).
• The Streaming Buffer (Kafka): Route these event payloads into dedicated Apache Kafka topics. This decoupling layer provides durable storage and acts as a shock absorber during massive transactional bursts on the source.
• The Ingestion Engine (ClickHouse): Provision a table using the Kafka engine within ClickHouse. This acts as an active consumer group, continuously pulling batches of raw events directly from the Kafka broker.
• State Materialization (ClickHouse MVs): Deploy a Materialized View to act as the transformation pipeline. It automatically reads from the Kafka engine table, applies necessary transformations, and routes the final data into a persistent storage table—most often a ReplacingMergeTree or CollapsingMergeTree to efficiently process mutations and deduplicate records.