EmergencyEMERGENCY? Get 24/7 Help Now!

Rescuing a crashed pt-online-schema-change with pt-archiver

 | June 30, 2016 |  Posted In: MySQL, Percona Server for MySQL, Percona Software, Percona Toolkit, Replication

PREVIOUS POST
NEXT POST

crashed pt-online-schema-changeThis article discusses how to salvage a crashed pt-online-schema-change by leveraging pt-archiver and executing queries to ensure that the data gets accurately migrated. I will show you how to continue the data copy process, and how to safely close out the pt-online-schema-change via manual operations such as RENAME TABLE and DROP TRIGGER commands. The normal process to recover from a crashed pt-online-schema-change is to drop the triggers on your original table and drop the new table created by the script. Then you would restart pt-online-schema-change. In this case, this wasn’t possible.

A customer recently needed to add a primary key column to a very busy table (with around 200 million rows). The table only had a unique key on one column (called our_id below). The customer had concerns about slave lag, and wanted to ensure there was little or no lag. This, as well as the fact that you can’t add a primary key as an online DDL in MySQL and Percona Server 5.6, meant the obvious answer was using pt-online-schema-change.

Due to the sensitivity of their environment, they could only afford one short window for the initial metadata locks, and needed to manually do the drop swap that pt-online-schema-change normally does automatically. This is where no-drop-triggers and no-swap-tables come in. The triggers will theoretically run indefinitely to keep the new and old tables in sync once pt-online-schema-change is complete. We crafted the following command:

You can see some of the specifics of other flags and why we used them in the Percona Toolkit Manual.

Once we ran the command the customer got concerned, as their monitoring tools weren’t showing any work done (which is by design, pt-online-schema-change doesn’t want to hurt your running environment). The customer ran strace -p to verify it was working. This wasn’t a great choice as it crashed pt-online-schema-change.

At this point, we knew that the application (and management) would not allow us to take new metadata locks to create triggers on the table, as we had passed our metadata lock window.

So how do we recover?

First, let’s start with a clean slate. We issued the following commands to create a new table, where __largetable_new is the table created by pt-online-schema-change:

Now the triggers on the original table, largetable are updating the new empty table that has our new schema.

Now let’s address the issue of actually moving the data that’s already in largetable to __largetable_new. This is where pt-archiver comes in. We crafted the following command:

We use pt-archiver to slowly copy records non-destructively to the new table based on our_id and WHERE 1=1 (all records). At this point, we periodically checked the MySQL data directory over the course of a day with ls -l to compare table sizes.

Once the table files were close to the same size, we ran counts on the tables. We noticed something interesting: the new table had thousands more records than the original table.

This concerned us. We wondered if our “hack” was a mistake. At this point we ran some verification queries:

We learned that there were older records that didn’t exist in the live table. This means that pt-archiver and the DELETE trigger may have missed each other (i.e., pt-archiver was already in a transaction but hadn’t written records to the new table until after the DELETE trigger already fired).

We verified with more queries:

They returned nothing.

Our result showed 4000 extra records in the new table. This shows that we ended up with extra records that were deleted from the original table. We ran other queries based on their data to verify as well.

This wasn’t a huge issue for our application, and it could have been easily dealt with using a simple DELETE query based on the unique index (i.e., if it doesn’t exist in the original table, delete it from the new one).

Now to complete the pt-online-schema-change actions. All we need to do is the atomic rename or drop swap. This should be done as soon as possible to avoid running in a degraded state, where all writes to the old table are duplicated on the new one.

Then drop the triggers for safety:

At this point it is safer to wait for the old table to clear out of the buffer pool before dropping it, just to ensure there is no impact on the server (maybe a week to be safe). You can check information_schema for a more accurate reading on this:

Once this goes to 0 you can issue:

PREVIOUS POST
NEXT POST
Manjot Singh

Manjot Singh is an Architect with Percona in California. He loves to learn about new technologies and apply them to real world problems. Manjot is a veteran of startup and Fortune 50 enterprise companies alike with a few years spent in government, education, and hospital IT.

One Comment

  • @manjot Singh, I am just curious to know pt-archiver copies the data in repeatable-read isolation level which is the reason for conflict with DELETE trigger ? Is it possible to mention isolation to be used for pt-online-schema-change as well as for pt-archiver ?

Leave a Reply