I’ve recently been trying to improve the performance of the Vertica replicator, particularly in the form of the of the new single schema replication. We’ve done a lot in the new Tungsten Replicator 5.3.0 release to improve (and ultimately support) the new single schema model.
As part of that, I’ve also been personally looking to Kodiak MemCloud as a deployment platform. The people at Kodiak have been really helpful (disclaimer: I’ve worked with some of them in the past). MemCloud is a high-performance cloud platform that is based on hardware with high speed (and volume) RAM, SSD and fast Ethernet connections. This means that even without any adjustment and tuning you’ve got a fast platform to work on.
However, if you are willing to put in some extra time, you can tune things further. Once you have a super quick environment, you find you can tweak and update the settings a little more because you have more options available. Ultimately you can then make use of that faster environment to stretch things a little bit further. And that’s exactly what I did when trying to determine how quickly I could get data into Vertica from MySQL.
In fact, the first time I ran my high-load test suite on MemCloud infrastructure, replicating data from MySQL into Vertica, I made this comment to my friend at Kodiak:
The whole thing went so quick I thought it hadn’t executed at all!
In general, there are two things you want to test when using replication to move data from a transactional environment into an analytical one:
- Latency of moving the data
- Apply rate for moving the data
The two are subtly different. The first measures how long it takes to get the data from the source to the target. The second measures how much data you can move in a set period of time.
Depending on your deployment and application, either, or both, can be critical. For example, if you are using analytics to perform real-time analysis and charging on your data, the first one is the most important, because you want the info up to date as quickly as possible. If you performing log analysis or longer-term trends, the second is probably more important. You may not worry about being a few seconds, but you want many thousands of transactions to be transferred. I concentrated on the former rather than the latter, because latency in the batch applier is something you can control by setting the batch interval.
So what did I test?
At a basic level, I was replicating data from MySQL directly into Vertica. That is, extracting data from the MySQL binary log, and writing that into a table within HPE Vertica cluster using 3 nodes. Each is running in MemCloud, which each running with 64GB of RAM and 2TB of SSD disk space. I’ve deliberately made no configuration changes to Vertica at this point.
The first thing I did was set-up a basic replication pipeline between the two. Replication into Vertica works by batch-loading data through CSV files into Vertica tables and then ‘materialising’ the changes into the carbon copy tables. Because it’s done in batches, the latency is effectively governed by the batch apply settings, which were configured for 10,000 rows or 5 seconds.
To generate the load, I’ve written a script that generates 100,000 rows of random data, then updates about 70% of those randomly and deletes the other 30%. So each schema is generating about 200,000 rows of changes for each load. This is designed to test the specific batch replication scenario. Ultimately it does this across multiple schemas (same structure). I specifically use this because I match this with the replication to get replication (rather than transaction) statistics. I need to be able to effectively monitor the apply rate into Vertica from MySQL.
The first time I ran the process of just generating the data and inserting into MySQL, the command returned almost immediately. I seriously thought it had failed because I couldn’t believe I’d just inserted 200,000 rows into MySQL that quick. Furthermore, over on the Vertica side, I’m monitoring the application through the trepctl perf command, which provides the live output of the process. And for a second I see the blip as the data is replicated and then applied. I thought it was so quick, it was a single row (or even failed transaction) that caused the blip.
The first time I ran the tests, I got some good results with 20 simultaneous schemas:
- 460,000 rows/minute from a single MySQL source into Vertica.
Then I doubled up the source MySQL servers, so two servers, 40 simultaneous schemas, and ultimately writing in about 8 million rows:
- 986,000 rows/minute into Vertica across 40 schemas from 2 sources
In both cases, the latency was between 3-7 seconds for each batch write (remember we are handling 10,000 rows or 5s per batch). We are also doing this across *different* schemas at this stage. These are not bad figures.
I did some further tweaks, this time, reconfiguring the batch writes to do larger blocks and larger intervals. This increases the potential latency (because there will be bigger gaps between writes into Vertica), but increases the overall row-apply performance. Now we are handling 100,000 rows or 10s intervals. The result? A small bump for a single source server:
- 710,000 rows/minute into Vertica across 20 schemas from 1 source
Latency has increased though, with us topping out at around 11.5s for the write when performing the very big blocks. Remember this is single-source, and so I know that the potential is there to basically double that with a second MySQL source since the scaling seems almost linear.
Now I wanted to move on to test a specific scenario I added into the applier, which is the ability to replicate from multiple source schemas into a single target schema. Each source is identical, and to ensure that the ‘materialise’ step works correctly, and that we can still analyse the data, a filter is inserted into the replication that adds the source schema to each row.
The sample data inserts look like this:
insert into msg values (0,"RWSAjXaQEtCf8nf5xhQqbeta"); insert into msg values (0,"4kSmbikgaeJfoZ6gLnkNbeta"); insert into msg values (0,"YSG4yeG1RI6oDW0ohG6xbeta");
With the filter, what gets inserted is;
0, "RWSAjXaQEtCf8nf5xhQqbeta", "sales1" 0,"4kSmbikgaeJfoZ6gLnkNbeta", "sales1"
Etc, where ‘sales1’ is the source schema, added as an extra column to each row.
This introduces two things we need to handle on the Vertica side:
- We now have to merge taking into account the source schema (since the ID column of the data could be the same across multiple schemas). For example, whereas before we did ‘DELETE WHERE ID IN (xxxx)’, and now we have to do ‘DELETE WHERE ID IN (xxxx) AND dbname = ‘sales1”.
- It increases the contention ratio on the Vertica table because we now effectively write into only one table. This increases the locks and the extents processed by Vertica.
The effect of this change is that the overall apply rate slows down slightly due to the increased contention on a single table. Same tests, 20 schemas from one MySQL source database and we get the following:
- 760,000 rows/minute into Vertica with a single target table
This is actually not as bad as I was expecting when you consider that we are modifying every row of incoming data, and are no longer able to multi-thread the apply.
I then tried increasing that using the two sources and 40 schemas into the same single table. Initially, the performance was no longer linear, and I failed to get any improvement beyond about 10% above that 760K/min figure.
Now it was time to tune other things. First of all, I changed some of the properties on the Vertica side in terms of the queries I was running, tweaking the selecting and query parameters for the DELETE operations. For batch loading, what we do is DELETE and then INSERT, or, in some case, DELETE and UPDATE if you’ve configured it that way. Tweaking the subquery that is being used increased the performance a little.
Changing the projections used also increased the performance of the single schema apply. But the biggest gains, perhaps unsurprisingly, were to change the way the tables were defined in the first place and to use partitions in the table definition. To do this, I modified the original filter that was adding the schema name, and instead had it add the schema hash, a unique Java ID for the string. Then I created the staging and base tables in Vertica using the integer hash as the partition. Then I modified the queries to ensure that the partitions would be used effectively.
The result was that the 760k/min rate was now scalable. I couldn’t get any faster when writing into a single schema, but the rate remains relatively constant whether I am replicating five schemas into a single one, or 40 schemas from two or three sources into the same. In fact, it does ultimately start to dip slightly as you add more source schemas.
Even better, the changes I’d made to the queries and the overall Vertica batch applier also improved the speed of the standard (i.e. multi-schema) applier. I also added partitions to the ID field for too to improve the general apply rate. After testing for a couple days, the average rate:
- 1,900,000 rows/minute from a single MySQL source into Vertica.
This was also scalable. Five MySQL sources elicited a rate of 8.8 million rows/minute into Vertica, making the applier rate linear with a 1% penalty for each additional MySQL source. The latency stayed the same, hovering around the same level as before of around 11s for most of the time. Occasionally you’d get a spike because Vertica was having trouble keeping up, but
Essentially, we are replicating data from MySQL into Vertica for analytics at a rate I simply called ‘outstandingly staggering’. I still do.
The new single schema applier, database name filter (rowadddbname) and performance improvements are all incorporated into the new Tungsten Replicator.