Google compute has taken some awesome strides in the last year. I now prefer GCE over AWS, because of pricing (per minute) and simplicity. AWS quickly can get overwhelming and overly complex.
Not just Compute, Google Network, Cloud Storage, Container Engine (Kubernetese is OpenSource, so you can run it internally too), Cloud Shell, Pub/Sub, their new Cloud Console, are awesome.
Pub/Sub is pretty meh. The Google Cloud team still don't seem to understand the need for a Kafka-like unified log service, unlike AWS (Kinesis is 2 years old this week) or even IBM Bluemix (who just launched Message Hub, which is true hosted Kafka).
One thing you gotta remember is, PubSub charges per volume, regardless of speed (in other words, scaling is free). AWS will charge you varying orders of magnitudes for varying scales, in addition to volume.
Cloud Pub/Sub is really a competitor to Amazon SQS, not Kinesis. It's more helpful to think of Kafka and Kinesis as databases containing first-class, immutable streams; writing to the streams and reading from the streams are completely decoupled, unlike in a traditional pub-sub system. Jay Kreps' blog post explains it better than I can:
In my limited Pub/Sub experience, this seems to be how it works. You publish to a topic (an immutable stream), and then create a decoupled subscription that reads messages from the topic. Am I missing something?
I think this sentence [1] helps to explain the difference:
> When you create a subscription, the system establishes a sync point. That is, your subscriber is guaranteed to receive any message published after this point.
With Kafka or Kinesis, I can write events to a stream/topic completely independently of any consumer. I can then bring as many consumers online as I want, and they can start processing from the beginning of my stream if they want. If one of my consumers has a bug in it, I can ask it to go back and start again. That's what I mean by an immutable stream in Kafka or Kinesis.
Cloud Pub/Sub engineer here. You can create as many consumers as you want. You can create them offline and bring them up and down whenever you want. Each consumer will receive a full copy of the stream, starting with its sync point (subscriber creation). Each message is delivered, and redelivered, to each consumer until that consumer acks that message.
If I understand your point correctly, the only expectation we haven't matched is the ability to "go back and start again". We hear you.
From your comment it sounds like you haven't used Kinesis or Kafka yourself - rather than take my word for it, I'd suggest your team give both of those platforms a serious try-out to really understand the capability gaps. I'd be surprised if a lot of your [prospective] customers weren't asking for these kinds of unified log capabilities in Cloud Pub/Sub.
Let me see if I'm understanding the criticism: when creating a consumer, the sync point of a new consumer really should start from the very beginning of the topic, at a predictable explicit start point, rather than at the current end of the topic. This makes a lot of sense, and yes, there is a disconnect between the models. We think the capabilities you are talking about are great and those use cases are important. All I can say is keep your eyes open.
We went with defaults from Google's internal use of Pub/Sub, which is older than the public release of Kinesis and Kafka. Internal use involves an approach where topics and consumers are very long-lived. Topics are high throughput, in terms of bytes published per unit time. Retaining all messages and starting consumers from the very beginning wasn't a sensible default; our focus was more centered on making sure that, once topics and consumers were set up, consumers could keep up over time.
In the work described by that video, they were essentially publishing tweets in real time into a Cloud Pub/Sub topic, thus making an "all tweets on Twitter in realtime" topic. This is a great example of a topic where producers and consumers are completely decoupled from each other. It doesn't necessarily make sense to retain all tweets forever by default (although there certainly are use cases for that). There are plenty of use cases where a consumer might want to say "ok, please start retaining all tweets made from here on out" rather than starting from a specific tweet.
> when creating a consumer, the sync point of a new consumer really should start from the very beginning of the topic, at a predictable explicit start point, rather than at the current end of the topic
I'll talk about Kinesis because that's the technology we use more at Snowplow. When creating a Kinesis consumer, I can specify whether I want to start reading from a) TRIM_HORIZON (which is the earliest events in the stream which haven't yet been expired aka "trimmed"), b) LATEST which is the Cloud Pub/Sub capability, c) AT_SEQUENCE_NUMBER {x} which means from the event in the stream with the given offset ID or d) AFTER_SEQUENCE_NUMBER {x} which is the event immediately after c).
Kinesis streams or Kafka topics don't themselves care about the progress of any individual consumer - consumers are responsible for tracking their own position in the stream via sequence numbers / offset IDs.
> It doesn't necessarily make sense to retain all tweets forever by default (although there certainly are use cases for that)
Completely agree. I think a good point of distinction between pub/sub systems and unified log is: use pub/sub when the messages are a means-to-an-end (which is feeding one or more downstream apps); use unified log when the events are an end-in-themselves (i.e. you would still want to preserve the events even if there were no consumers live).
Anyway, I could talk about this stuff all day :-) - if you'd like to chat further, my details are in my profile!
I haven't played with kafka in a while, but basically,
1. each group id represents a point in the stream that a consumer is processing off of. You could technically have multiple processes consuming off of a single group id.
2. there was a configuration on time to keep things there as well as space if I remember correctly, but basically, there has to be. There's a pretty hard limit on what all you can store on disk.
edit: changed consumer id to group id. If you want more info, feel free to ping me about the ecosystem
Pricing is fine. Tech support is a joke. I wish I could pay more for Google to actually offer good support. We tried GCE for a while but it just didn't cut it in terms of mean response time to questions. AWS really leads on support.
Well tbh we got answers but they usually were; nope not offered, cannot be fixed. 3 weeks later I met a Google engineer working with GCE at a party who fixed it for us the day after hearing about our issue, which means it probably was able to be fixed from the first time.
I haven't used GCE but I've had one of the worst support experiences ever with Google Cloud Billing Support while using Google App Engine.
I had a very simple question about billing (why was my bill higher than it seemed like it should be). Each reply would take a week and they'd often consist of copy and paste messages asking me to enter information I had already supplied or requests that I take screenshots of my console (all information they should readily have available). Then right at the end they swapped out who I was talking to for someone else and asked me to look up more information they already had and just ignored my question in the last email which would mean another week before I got an answer.
I had, luckily, experimented with the configuration and figured out what was wrong. The default instance class in reality is one higher (F2) than the documentation says it is ("If you do not specify a class, F1 is assigned by default."). Nowhere on the Console does it list what instance class is being used (which would have made the problem obvious) so there was really no way of knowing this without just guessing what the problem was. They never did answer my question "What is the default instance class?" (instead just abruptly ended the support ticket after proposed my theory about what was wrong).
Then I started getting emails about a billing account being past due. It was an old billing account from before I moved to High Replication (I have no idea how I ended up with two billing accounts...it was during the dark time when the console was even worse than it is now). That billing account was assigned to no projects and had no oustanding balance. I jumped in and just deleted the unused billing account. Then a few days later they sent a scary email saying that the billing account had been terminated (even though I had deleted it) which made my scramble to make sure they didn't close my in-use billing account out of nowhere (they hadn't thankfully).
None of this has left me with any confidence in Google's Cloud offerings.
I plan to migrate off GAE as soon as I can rewrite the app (luckily it's not very big).
Can you say more about your experience with AWS support? My experience has been consistently quite bad. I've had to resort to AWS technical support perhaps a dozen times over the last few years. They're always slow to respond (even if you have an SLA promising otherwise), and it always takes half a dozen go-rounds to get to a resolution. They will stick with you until the issue is resolved, but they don't add much value along the way. It's a slow process of "did you do obvious thing X" (even if you've already provided facts indicating that couldn't be the problem), "please provide six different pieces of information" (that don't seem to bear on the actual question), and "I've never seen this before, let me go research it".
For the lowest support plan (Silver), high priority issues were addressed pretty quickly. Low priority ones are a bit slow, but hey, they're low priority. Plus the Silver support plan is very cheap.
Bigtable is best thought of as an "event database". High reads, high writes, single index, accessible through the Hbase API. Cassandra and Hbase are similar technologies that are inspired by the original Bigtable paper.
One big benefit of Bigtable is its scalability. To scale up, you turn the 'scale' knob. By contrast, Cassandra and Hbase are headaches to scale (Apple has acquired Cassandra companies to aid in operation and scale).
Here's a couple of guys from Sungard, who scaled to about 3,000,000 writes per second with a couple weekends' worth of effort (something only few beyond the likes of Facebook, Netflix, and Apple can achieve) https://cloud.google.com/bigtable/pdf/SunGardCATCaseStudy.pd...
Hey. I'm one of the "guys from SunGard", although I'm no longer there. The longer version is this: https://cloud.google.com/bigtable/pdf/ConsolidatedAuditTrail... . A lot of it is related to the use case, but yeah, Bigtable handled pretty much whatever we wanted to throw at it. No other cloud provider can offer this sort of scale and performance right now without a ton of manual management or significant compromises, something that seems to have yet to sink in (although few companies need the scale we went up to).
It did take a lot more work than "a couple weekends" though :).
