Great news from influxdata. Though all of these features (clustering, monitoring, and data elxploration) could be released with standard edition. The start of InfluxDB looked very promising, not sure if people feel the same with the chosen marketing model.
Yes, and there are many ways to achieve the goal. Some of them are listed in "pandemicsyn" comment. I still think, that the plans, especially with clustering side should have been announced earlier, instead of starting it as an open source, attracting audience, then dropping and creating just enterprise version.
Honestly, this is a very fine line to balance. If you open source your core technology, a lot of those "services" you listed are open to competition from other companies with little to no barrier to entry.
There is a lot of discussion going on currently around OSS business models. The general consensus is that support contracts / professional services / support do not monetize well enough to grow a large business.
Hosted services is a viable business model. The risk here is to ensure that your product, if successful, cannot simply be deployed by Amazon. Dual licensing has been one response to the Amazon risk.
Admin tools alone are also thought to be difficult to monetize at sufficient value to grow a large business.
While DataStax has a nice UI, they have built a data platform with Cassandra the core, much like Cloudera has built a data platform with HDFS/HBase at the core. Personally, I view the data platforms as prepackaged and productized SI (which is very valuable..
I think some people feel disgruntled because the development of clustering started as open source and was some way down the road to being production ready, then it was removed and moved to the enterprise version only. I was planning a metrics architecture around influx; but following that announcement, influx was out of the picture. Had I known from the start that clustering was going to be enterprise only, likely I still wouldn't have chosen it, but my takeaway thought would have been "wow, this is cool, shame clustering is enterprise only". Now I just feel a mild sense of deflation.
I don't blame them for going down this route (we all have to eat), but I think had that descision been made earlier it would have looked better on them.
I really like the influxdb interface -- reading and writing data are super simple, but in some recent testing I did, it really really hated high-cardinality data sets.
After 1.0 is released our two highest priorities are to handle high cardinality data sets (on disk and in memory inverted index) and continuous queries that are more automatic (downsample everything and give users the option for query results that scale automatically like Graphite)
Right on -- for the record, this data is ending up in mongodb at the moment, so I need to write logic for binning by time and so forth myself. Our use case is maybe odd in this space -- sampling a handful of values per item a couple times a day, but for millions of items.
I compared a dataset to postgres and the new mongo storage engine is very good -- dunno if wiredTiger is something you guys are looking at, but it's very compact and fast to query with the proper indexes.
i had a similar issue with mongo and time series data. for special purpose event data like that gist, i think you should consider using a column store.
We evaluated InfluxDB for our TSDB solution and found it to be generally high performance (up to about 100k unique series). Kapacitor also had some promising features for realtime aggregations, but was ultimately pretty buggy and didn't have support for out-of-order event processing. All in all, it looks like a great suite of products and we're curious to see how it evolves over the next few years.
Like many other users here, we were disappointed about paid clustering, but when the original press release said $400, we were willing to wait it out and see. However, we ultimately decided to go a different direction after seeing they wanted $20k+ to run clustering on a 256GB node. We ingest 10s of millions of data points per day from IoT sensors, and expect our data size to far exceed that capacity.
That said, we plan to run our own Cassandra cluster w/ KairosDB (http://kairosdb.github.io/) acting as a read / write abstraction layer. It'll cost us about $11k to run the cluster for the year, with 3 nodes @300GB/ea., leveraging Cassandra's (free) and open source clustering, HA, and replication technology.
+1 for Cassandra and KairosDB. We use Instaclustr to host Cassandra on AWS and do time series data management with Kairos. It's been a hugely scalable, robust system we use ingest about 5B datapoints per day.
Also, Brian (project leader of KairosDB) agreed to a small consulting project for us to help us configure and tune Kairos. He's a very nice guy and was very helpful.
I have just shut down 95% of my collectd/graphite infrastructure after migrating it over to InfluxDB+telegraf+grafana. I'm loving it! Since shutting down collectd, system load and io wait time across my fleet has gone way down, and available CPU has gone way up.
Though I wouldn't say it was a smooth transition. I started with 0.8, IIRC, and while it worked ok it used an amazing amount of storage. 4GB for a year worth of graphite data blew up to 100GB for a month of InfluxDB.
I gave up on InfluxDB a few times during the process, but at 0.11 I tried it again and is has been pretty good. We are only putting the Telegraf data and one small service statistic in it, but the storage is pretty reasonable at 12GB for a few months of data. Querying and graphing the data with Grafana is great.
If you have tried it before 0.11, definitely try it again. The guys giving a Prometheus talk at PyCon were really down on InfluxDB, but they hadn't tried it for 6 months. I was like "Yeah, it was unusable then". I wanted to like Prometheus, but I just couldn't figure out how to feed my data into it.
Have you tried 0.13? How much disk space is it using? It's supposed to be "better than graphite".
Really curious as I am looking forward to set up InfluxDB (moving from graphite too). I was going to use collectd, but your comment makes me wonder if it's the right choice.
It's really hard to compare and say better or worse than graphite, because it isn't an apples/apples comparison. I, so far, haven't figured out how to do roll up like graphite has built in, so I think I'm holding onto the high precision data. So that makes it not even close to a fair comparison. Still though, 12GB for our current data set feels reasonable. Back in 0.8 when it was more like 150GB for a month, that was not gonna work.
For a while I was feeding collectd into InfluxDB. It didn't really produce anything I could use immediately, and I haven't gone back to revisit it.
18 comments
[ 4.3 ms ] story [ 47.8 ms ] threadHosted services is a viable business model. The risk here is to ensure that your product, if successful, cannot simply be deployed by Amazon. Dual licensing has been one response to the Amazon risk.
Admin tools alone are also thought to be difficult to monetize at sufficient value to grow a large business.
While DataStax has a nice UI, they have built a data platform with Cassandra the core, much like Cloudera has built a data platform with HDFS/HBase at the core. Personally, I view the data platforms as prepackaged and productized SI (which is very valuable..
I don't blame them for going down this route (we all have to eat), but I think had that descision been made earlier it would have looked better on them.
https://docs.influxdata.com/influxdb/v0.10/guides/hardware_s...
Show stopper if you're in the same boat that I am.
I compared a dataset to postgres and the new mongo storage engine is very good -- dunno if wiredTiger is something you guys are looking at, but it's very compact and fast to query with the proper indexes.
Dumped some stats into a gist as well: https://gist.github.com/mattbillenstein/89969980025414e2bca8...
Like many other users here, we were disappointed about paid clustering, but when the original press release said $400, we were willing to wait it out and see. However, we ultimately decided to go a different direction after seeing they wanted $20k+ to run clustering on a 256GB node. We ingest 10s of millions of data points per day from IoT sensors, and expect our data size to far exceed that capacity.
That said, we plan to run our own Cassandra cluster w/ KairosDB (http://kairosdb.github.io/) acting as a read / write abstraction layer. It'll cost us about $11k to run the cluster for the year, with 3 nodes @300GB/ea., leveraging Cassandra's (free) and open source clustering, HA, and replication technology.
Also, Brian (project leader of KairosDB) agreed to a small consulting project for us to help us configure and tune Kairos. He's a very nice guy and was very helpful.
Though I wouldn't say it was a smooth transition. I started with 0.8, IIRC, and while it worked ok it used an amazing amount of storage. 4GB for a year worth of graphite data blew up to 100GB for a month of InfluxDB.
I gave up on InfluxDB a few times during the process, but at 0.11 I tried it again and is has been pretty good. We are only putting the Telegraf data and one small service statistic in it, but the storage is pretty reasonable at 12GB for a few months of data. Querying and graphing the data with Grafana is great.
If you have tried it before 0.11, definitely try it again. The guys giving a Prometheus talk at PyCon were really down on InfluxDB, but they hadn't tried it for 6 months. I was like "Yeah, it was unusable then". I wanted to like Prometheus, but I just couldn't figure out how to feed my data into it.
Really curious as I am looking forward to set up InfluxDB (moving from graphite too). I was going to use collectd, but your comment makes me wonder if it's the right choice.
It's really hard to compare and say better or worse than graphite, because it isn't an apples/apples comparison. I, so far, haven't figured out how to do roll up like graphite has built in, so I think I'm holding onto the high precision data. So that makes it not even close to a fair comparison. Still though, 12GB for our current data set feels reasonable. Back in 0.8 when it was more like 150GB for a month, that was not gonna work.
For a while I was feeding collectd into InfluxDB. It didn't really produce anything I could use immediately, and I haven't gone back to revisit it.