pgwarehouse - quickly sync Postgres data to your cloud warehouse

Introduction

Postgres is an amazing, general purpose OLTP database. But it's not designed for heavy analytic (OLAP) usage. Analytic queries are much better served by a columnar store database like Snowflake or Clickhouse.

This package allows you to easily sync data from a Postgres database into a local or cloud data warehouse (currently Snowflake or Clickhouse). You can perform a one-time sync operation, or run periodic incremental syncs to keep your warehouse up to date.

Features

• High performance by using COPY to move lots of data efficiently. pgwarehouse can easily sync hundreds of millions of rows of data (tens of GB) per hour.
• Supports multiple update strategies for immutable or mutable tables.
• Easy to configure and run.

Installation

pip install pgwarehouse

Now you need to configure credentials for your Postgres source and the warehouse destination.

You can place Postgres credentials either in your config file or in your environment. If using the environment you need to set these variables:

PGHOST
PGDATABASE
PGUSER
PGPASSWORD
PGSCHEMA (defaults to 'public')

Creating a config file

Run this command to create a template config file:

pgwarehouse init

This will create a local pgwarehouse_conf.yaml file. Now you can edit your Postgres credentials in the postgres stanza of the config file:

postgres:
    pghost: (defaults to $PGHOST)
    pgdatabase: (defaults to $PGDATABASE
    pguser: (defaults to $PGUSER)
    pgpassword: (defaults to $PGPASSWORD)
    pgschema: (defaults to 'public')

Specifying the warehouse credentials

Again you can use the environment or the config file. Set these sets of vars in your env:

CLICKHOUSE_HOST
CLICKHOUSE_DATABASE
CLICKHOUSE_USER
CLICKHOUSE_PWD

or

SNOWSQL_ACCOUNT
SNOWSQL_DATABASE
SNOWSQL_SCHEMA
SNOWSQL_WAREHOUSE
SNOWSQL_USER
SNOWSQL_PWD

(The Snowflake parameters are the same as those for the SnowSQL CLI tool. The SNOWSQL_ACCOUNT value should be your "account identifier".)

or set these values in the warehouse stanza in the config file:

warehouse:
    backend: (clickhouse|snowflake)
    clickhouse_host: 
    clickhouse_database: 
    clickhouse_user:
    clickhouse_password:
    --or--
    snowsql_account:
    snowsql_database:
    snowsql_schema:
    snowsql_warehouse:
    snowsql_user:
    snowsql_pwd:

Usage

Once the credentials are configured you can start syncing data. Start by listing tables from the Postgres database:

pgwarehouse list

And you can see which tables exist so far in the warehouse:

pgwarehouse listwh

Now use sync to sync a table (eg. the 'users' table):

pgwarehouse sync users

Data will be downloaded from the Postgres database into CSV files on the local machine, and then those files will be uploaded to the warehouse. Running pgwarehouse listwh will show the new table.

Updating a table

After the initial sync has run, you can update the warehouse table with new records by running sync again:

pgwarehouse sync users

See update strategies for different ways to update your table on each sync.

Syncing multiple tables

There are two ways to manage multiple tables. The first is just to pass all in place of the table name:

pgwarehouse sync all

This will attempt to sync ALL tables from Postgres into the warehouse. This could take a while!

The other way is to specify the tables list in the config file:

tables:
    - users
    - charges
    - logs

Now when you specify sync all the tool will use the list of tables specified in the config file.

Pro tip! You can add the max_records settings to your postgres configuration to limit the number of records copied per table. This can be useful for testing the initial sync in case you have some large tables. Set this value to something reasonable (like 10000) and then try syncrhonizing all tables to make sure they copy properly. Once you have verified the tables in the warehouse then you can remove this setting, drop any large tables, and then copy them in full (just run sync all again).

Table update strategies

New Records Only (default)

The default update strategy is "new records only". This is done by selecting records with a greater value for their primary id column than the greatest value currently in the warehouse. This strategy is simple and quick, but only works for monotonically incrementing primary keys, and only finds new records.

Reload each time

Another supported strategy is "reload each time". This is the simplest strategy and we simply reload the entire table every time we sync. This strategy should be fine for small-ish tables (like <10m rows).

Last Modified

Finally, if your table has a last modified column then you can use the "all modifications strategy". In this case all records with a last modified timestamp greater than the maximum value found in the warehouse will be selected and "upserted" into the warehouse. Records that are already present (via matching the primary key) will be updated, and new records will be inserted.

• The Snowflake backend uses the MERGE operation.
• The Clickhouse backend uses ALTER TABLE .. DELETE to remove matching records and then INSERT to insert the new values.

What about deletes?

There is no simple way to capture deletes - you have to reload the entire table. A common pattern is to apply new records on a daily basis, and reload the entire table every week to remove deleted records.

What if my table has no primary key?

All the update strategies except "reload each time" require your table to have a primary key column.

Specifying update strategy at the command line

pgwarehouse sync <table>   (defaults to NEW RECORDS)
pgwarehouse sync <table> last_modified=<last modified column>   (MODIFIED RECORDS)
pgwarehouse reload <table> (reloads the whole table)

Specifying update strategy in the config file

You can configure the update strategy selectively for each table in the config file. To do so, specify the table as a nested dictionary with options:

tables:
    - accounts
    - users:
        reload: true
    - orders:
        last_modified: updated_at
    - shoppers
        last_modified: update_time
        reload: sun
    - original_orders:
        skip: true

In this example:

• accounts will have new records only applied at each sync
• users will be reloaded completely on each sync
• orders will have modified records (found by the 'updated_at' column) applied on each sync
• shoppers will have modified records applied on each sync, except for any sync which happens on Sunday, in which case the entire table will be reloaded.
• original_orders will be skipped entirely

The reload argument can take 3 forms:

reload: true    - reload the table every sync
reload: [sun,mon,tue,wed,thur,fri]  - reload if the sync occurs on this day of the week
reload: 1-31    - reload if the sync occurs on this numeric day of the month (don't use 31!)

Scheduling regular data syncs

pgwarehouse does not including any scheduling itself, you will need an external trigger like cron, Heroku Scheduler, or a K8s CronJob.

When running, the tool will need access to local storage - potentially a lot if you are synchronizing big tables. But nothing needs to persist between sync runs (except the config file) - the tool only relies on state it can query from Postgres or the warehouse.

Troubleshooting

Sometimes when you are testing things out it can be helpful to do the sync in two phases: 1)download the data, 2)upload the data. You can use extract and load for this:

pgwarehouse extract <table>     - only downloads data
pgwarehouse load <table>        - loads the data into the warehouse

When the extract process runs, its stores data in ./pgw_data/<table name>_data. As files are uploaded they are moved into an archive subdirectory. When the next sync runs then this archive directory will be cleaned up. This allows you to go examine the CSV downloaded data in case the upload fails for some reason.

Limitations

Column type mapping today is very limited. More esoteric column types like JSON or ARRAY are simply mapped as VARCHAR columns. Some of these types are supported in the warehouse and could be implemented more accurately.

Composite primary keys (using multiple columns) have limited support. Today they will only work with the RELOAD strategy.

Non-numeric primary key types (like UUIDs) probably won't work unless they have a good lexigraphic sort that supports a > where clause.