Hi Harsha, Thank you for proposing this KIP. We are looking forward to this feature as well.
A few questions around the design & implementation: 1. Wouldn't implicit checking for old offsets in remote location if not found locally on the leader i.e. do we really need remote index files? Since the storage path for a given topic would presumably be constant across all the brokers, the remote topic-partition path could simply be checked to see if there are any segment file names that would meet the offset requirements for a Consumer Fetch Request. RSM implementations could optionally cache this information. 2. Would it make sense to create an internal compacted Kafka topic to publish & record remote segment information? This would enable the followers to get updates about new segments rather than running list() operations on remote storage to detect new segments which may be expensive. 3. For RLM to scan local segment rotations are you thinking of leveraging java.nio.file.WatchService or simply running listFiles() on a periodic basis? Since WatchService implementation is heavily OS dependent it might create some complications around missing FS Events. Thanks. Ambud On Thu, Mar 28, 2019 at 8:04 AM Ron Dagostino <rndg...@gmail.com> wrote: > Hi Harsha. I'm excited about this potential feature. Did you consider > storing the information about the remote segments in a Kafka topic as > opposed to in the remote storage itself? The topic would need infinite > retention (or it would need to be compacted) so as not to itself be sent to > cold storage, but assuming that topic would fit on local disk for all time > (an open question as to whether this is acceptable or not) it feels like > the most natural way to communicate information among brokers -- more > natural than having them poll the remote storage systems, at least. > > To add to Eric's question/confusion about where logic lives (RLM vs. RSM), > I think it would be helpful to explicitly identify in the KIP that the RLM > delegates to the RSM since the RSM is part of the public API and is the > pluggable piece. For example, instead of saying "RLM will ship the log > segment files that are older than a configurable time to remote storage" I > think it would be better to say "RLM identifies log segment files that are > older than a configurable time and delegates to the configured RSM to ship > them to remote storage" (or something like that -- just make it clear that > the RLM is delegating to the configured RSM). > > Ron > > > > > > On Thu, Mar 28, 2019 at 6:12 AM Eno Thereska <eno.there...@gmail.com> > wrote: > > > Thanks Harsha, > > > > A couple of comments: > > > > Performance & durability > > ---------------------------------- > > - would be good to have more discussion on performance implications of > > tiering. Copying the data from the local storage to the remote storage is > > going to be expensive in terms of network bandwidth and will affect > > foreground traffic to Kafka potentially reducing its throughput and > > latency. > > - throttling the copying of the data above might be a solution, however > if > > you have a few TB of data to move to the slower remote tier the risk is > > that the movement will never complete on time under high Kafka load. Do > we > > need a scheduler to use idle time to do the copying? > > - Have you considered having two options: 1) a slow tier only (e.g., all > > the data on HDFS) and 2) a fast tier only like Kafka today. This would > > avoid copying data between the tiers. Customers that can tolerate a > slower > > tier with a better price/GB can just choose option (1). Would be good to > > put in Alternatives considered. > > > > Topic configs > > ------------------ > > - related to performance but also availability, we need to discuss the > > replication mode for the remote tier. For example, if the Kafka topics > used > > to have 3-way replication, will they continue to have 3-way replication > on > > the remote tier? Will the user configure that replication? In S3 for > > example, one can choose from different S3 tiers like STD or SIA, but > there > > is no direct control over the replication factor like in Kafka. > > - how will security and ACLs be configured for the remote tier. E.g., if > > user A does not have access to a Kafka topic, when that topic is moved to > > S3 or HDFS there needs to be a way to prevent access to the S3 bucket for > > that user. This might be outside the scope of this KIP but would be good > to > > discuss first. > > > > That's it for now, thanks > > Eno > > > > > > On Wed, Mar 27, 2019 at 4:40 PM Harsha <ka...@harsha.io> wrote: > > > > > Hi All, > > > Thanks for your initial feedback. We updated the KIP. Please > > > take a look and let us know if you have any questions. > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-405%3A+Kafka+Tiered+Storage > > > > > > Thanks, > > > Harsha > > > > > > On Wed, Feb 6, 2019, at 10:30 AM, Harsha wrote: > > > > Thanks Eno, Adam & Satish for you review and questions. I'll address > > > > these in KIP and update the thread here. > > > > > > > > Thanks, > > > > Harsha > > > > > > > > On Wed, Feb 6, 2019, at 7:09 AM, Satish Duggana wrote: > > > > > Thanks, Harsha for the KIP. It is a good start for tiered storage > in > > > > > Kafka. I have a few comments/questions. > > > > > > > > > > It may be good to have a configuration to keep the number of local > > > > > segments instead of keeping only the active segment. This config > can > > > > > be exposed at cluster and topic levels with default value as 1. In > > > > > some use cases, few consumers may lag over one segment, it will be > > > > > better to serve from local storage instead of remote storage. > > > > > > > > > > It may be better to keep “remote.log.storage.enable” and respective > > > > > configuration at topic level along with cluster level. It will be > > > > > helpful in environments where few topics are configured with > > > > > local-storage and other topics are configured with remote storage. > > > > > > > > > > Each topic-partition leader pushes its log segments with respective > > > > > index files to remote whenever active log rolls over, it updates > the > > > > > remote log index file for the respective remote log segment. The > > > > > second option is to add offset index files also for each segment. > It > > > > > can serve consumer fetch requests for old segments from local log > > > > > segment instead of serving directly from the remote log which may > > > > > cause high latencies. There can be different strategies in when the > > > > > remote segment is copied to a local segment. > > > > > > > > > > What is “remote.log.manager.scheduler.interval.ms” config about? > > > > > > > > > > How do followers sync RemoteLogSegmentIndex files? Do they request > > > > > from leader replica? This looks to be important as the failed over > > > > > leader should have RemoteLogSegmentIndex updated and ready to avoid > > > > > high latencies in serving old data stored in remote logs. > > > > > > > > > > Thanks, > > > > > Satish. > > > > > > > > > > On Tue, Feb 5, 2019 at 10:42 PM Ryanne Dolan < > ryannedo...@gmail.com> > > > wrote: > > > > > > > > > > > > Thanks Harsha, makes sense. > > > > > > > > > > > > Ryanne > > > > > > > > > > > > On Mon, Feb 4, 2019 at 5:53 PM Harsha Chintalapani < > > ka...@harsha.io> > > > wrote: > > > > > > > > > > > > > "I think you are saying that this enables additional > (potentially > > > cheaper) > > > > > > > storage options without *requiring* an existing ETL pipeline. “ > > > > > > > Yes. > > > > > > > > > > > > > > " But it's not really a replacement for the sort of pipelines > > > people build > > > > > > > with Connect, Gobblin etc.” > > > > > > > > > > > > > > It is not. But also making an assumption that everyone runs > these > > > > > > > pipelines for storing raw Kafka data into HDFS or S3 is also > > wrong > > > > > > > assumption. > > > > > > > The aim of this KIP is to provide tiered storage as whole > package > > > not > > > > > > > asking users to ship the data on their own using existing ETL, > > > which means > > > > > > > running a consumer and maintaining those pipelines. > > > > > > > > > > > > > > " My point was that, if you are already offloading records in > an > > > ETL > > > > > > > pipeline, why do you need a new pipeline built into the broker > to > > > ship the > > > > > > > same data to the same place?” > > > > > > > > > > > > > > As you said its ETL pipeline, which means users of these > > pipelines > > > are > > > > > > > reading the data from broker and transforming its state and > > > storing it > > > > > > > somewhere. > > > > > > > The point of this KIP is store log segments as it is without > > > changing > > > > > > > their structure so that we can use the existing offset > mechanisms > > > to look > > > > > > > it up when the consumer needs to read old data. When you do > load > > > it via > > > > > > > your existing pipelines you are reading the topic as a whole , > > > which > > > > > > > doesn’t guarantee that you’ll produce this data back into HDFS > in > > > S3 in the > > > > > > > same order and who is going to generate the Index files again. > > > > > > > > > > > > > > > > > > > > > "So you'd end up with one of 1)cold segments are only useful to > > > Kafka; 2) > > > > > > > you have the same data written to HDFS/etc twice, once for > Kafka > > > and once > > > > > > > for everything else, in two separate formats” > > > > > > > > > > > > > > You are talking two different use cases. If someone is storing > > raw > > > data > > > > > > > out of Kafka for long term access. > > > > > > > By storing the data as it is in HDFS though Kafka will solve > this > > > issue. > > > > > > > They do not need to run another pipe-line to ship these logs. > > > > > > > > > > > > > > If they are running pipelines to store in HDFS in a different > > > format, > > > > > > > thats a different use case. May be they are transforming Kafka > > > logs to ORC > > > > > > > so that they can query through Hive. Once you transform the > log > > > segment it > > > > > > > does loose its ability to use the existing offset index. > > > > > > > Main objective here not to change the existing protocol and > still > > > be able > > > > > > > to write and read logs from remote storage. > > > > > > > > > > > > > > > > > > > > > -Harsha > > > > > > > > > > > > > > On Feb 4, 2019, 2:53 PM -0800, Ryanne Dolan < > > ryannedo...@gmail.com > > > >, > > > > > > > wrote: > > > > > > > > Thanks Harsha, makes sense for the most part. > > > > > > > > > > > > > > > > > tiered storage is to get away from this and make this > > > transparent to > > > > > > > the > > > > > > > > user > > > > > > > > > > > > > > > > I think you are saying that this enables additional > > (potentially > > > cheaper) > > > > > > > > storage options without *requiring* an existing ETL pipeline. > > > But it's > > > > > > > not > > > > > > > > really a replacement for the sort of pipelines people build > > with > > > Connect, > > > > > > > > Gobblin etc. My point was that, if you are already offloading > > > records in > > > > > > > an > > > > > > > > ETL pipeline, why do you need a new pipeline built into the > > > broker to > > > > > > > ship > > > > > > > > the same data to the same place? I think in most cases this > > will > > > be an > > > > > > > > additional pipeline, not a replacement, because the segments > > > written to > > > > > > > > cold storage won't be useful outside Kafka. So you'd end up > > with > > > one of > > > > > > > 1) > > > > > > > > cold segments are only useful to Kafka; 2) you have the same > > > data written > > > > > > > > to HDFS/etc twice, once for Kafka and once for everything > else, > > > in two > > > > > > > > separate formats; 3) you use your existing ETL pipeline and > > read > > > cold > > > > > > > data > > > > > > > > directly. > > > > > > > > > > > > > > > > To me, an ideal solution would let me spool segments from > Kafka > > > to any > > > > > > > sink > > > > > > > > I would like, and then let Kafka clients seamlessly access > that > > > cold > > > > > > > data. > > > > > > > > Today I can do that in the client, but ideally the broker > would > > > do it for > > > > > > > > me via some HDFS/Hive/S3 plugin. The KIP seems to accomplish > > > that -- just > > > > > > > > without leveraging anything I've currently got in place. > > > > > > > > > > > > > > > > Ryanne > > > > > > > > > > > > > > > > On Mon, Feb 4, 2019 at 3:34 PM Harsha <ka...@harsha.io> > wrote: > > > > > > > > > > > > > > > > > Hi Eric, > > > > > > > > > Thanks for your questions. Answers are in-line > > > > > > > > > > > > > > > > > > "The high-level design seems to indicate that all of the > > logic > > > for > > > > > > > when and > > > > > > > > > how to copy log segments to remote storage lives in the RLM > > > class. The > > > > > > > > > default implementation is then HDFS specific with > additional > > > > > > > > > implementations being left to the community. This seems > like > > > it would > > > > > > > > > require anyone implementing a new RLM to also re-implement > > the > > > logic > > > > > > > for > > > > > > > > > when to ship data to remote storage." > > > > > > > > > > > > > > > > > > RLM will be responsible for shipping log segments and it > will > > > decide > > > > > > > when > > > > > > > > > a log segment is ready to be shipped over. > > > > > > > > > Once a Log Segement(s) are identified as rolled over, RLM > > will > > > delegate > > > > > > > > > this responsibility to a pluggable remote storage > > > implementation. > > > > > > > Users who > > > > > > > > > are looking add their own implementation to enable other > > > storages all > > > > > > > they > > > > > > > > > need to do is to implement the copy and read mechanisms and > > > not to > > > > > > > > > re-implement RLM itself. > > > > > > > > > > > > > > > > > > > > > > > > > > > "Would it not be better for the Remote Log Manager > > > implementation to be > > > > > > > > > non-configurable, and instead have an interface for the > > remote > > > storage > > > > > > > > > layer? That way the "when" of the logic is consistent > across > > > all > > > > > > > > > implementations and it's only a matter of "how," similar to > > > how the > > > > > > > Streams > > > > > > > > > StateStores are managed." > > > > > > > > > > > > > > > > > > It's possible that we can RLM non-configurable. But for the > > > initial > > > > > > > > > release and to keep the backward compatibility > > > > > > > > > we want to make this configurable and for any users who > might > > > not be > > > > > > > > > interested in having the LogSegments shipped to remote, > they > > > don't > > > > > > > need to > > > > > > > > > worry about this. > > > > > > > > > > > > > > > > > > > > > > > > > > > Hi Ryanne, > > > > > > > > > Thanks for your questions. > > > > > > > > > > > > > > > > > > "How could this be used to leverage fast key-value stores, > > e.g. > > > > > > > Couchbase, > > > > > > > > > which can serve individual records but maybe not entire > > > segments? Or > > > > > > > is the > > > > > > > > > idea to only support writing and fetching entire segments? > > > Would it > > > > > > > make > > > > > > > > > sense to support both?" > > > > > > > > > > > > > > > > > > LogSegment once its rolled over are immutable objects and > we > > > want to > > > > > > > keep > > > > > > > > > the current structure of LogSegments and corresponding > Index > > > files. It > > > > > > > will > > > > > > > > > be easy to copy the whole segment as it is, instead of > > > re-reading each > > > > > > > file > > > > > > > > > and use a key/value store. > > > > > > > > > > > > > > > > > > " > > > > > > > > > - Instead of defining a new interface and/or mechanism to > ETL > > > segment > > > > > > > files > > > > > > > > > from brokers to cold storage, can we just leverage Kafka > > > itself? In > > > > > > > > > particular, we can already ETL records to HDFS via Kafka > > > Connect, > > > > > > > Gobblin > > > > > > > > > etc -- we really just need a way for brokers to read these > > > records > > > > > > > back. > > > > > > > > > I'm wondering whether the new API could be limited to the > > > fetch, and > > > > > > > then > > > > > > > > > existing ETL pipelines could be more easily leveraged. For > > > example, if > > > > > > > you > > > > > > > > > already have an ETL pipeline from Kafka to HDFS, you could > > > leave that > > > > > > > in > > > > > > > > > place and just tell Kafka how to read these > records/segments > > > from cold > > > > > > > > > storage when necessary." > > > > > > > > > > > > > > > > > > This is pretty much what everyone does and it has the > > > additional > > > > > > > overhead > > > > > > > > > of keeping these pipelines operating and monitoring. > > > > > > > > > What's proposed in the KIP is not ETL. It's just looking a > > the > > > logs > > > > > > > that > > > > > > > > > are written and rolled over to copy the file as it is. > > > > > > > > > Each new topic needs to be added (sure we can do so via > > > wildcard or > > > > > > > > > another mechanism) but new topics need to be onboard to > ship > > > the data > > > > > > > into > > > > > > > > > remote storage through a traditional ETL pipeline. > > > > > > > > > Once the data lands somewhere like HDFS/HIVE etc.. Users > need > > > to write > > > > > > > > > another processing line to re-process this data similar to > > how > > > they are > > > > > > > > > doing it in their Stream processing pipelines. Tiered > storage > > > is to get > > > > > > > > > away from this and make this transparent to the user. They > > > don't need > > > > > > > to > > > > > > > > > run another ETL process to ship the logs. > > > > > > > > > > > > > > > > > > "I'm wondering if we could just add support for loading > > > segments from > > > > > > > > > remote URIs instead of from file, i.e. via plugins for > s3://, > > > hdfs:// > > > > > > > etc. > > > > > > > > > I suspect less broker logic would change in that case -- > the > > > broker > > > > > > > > > wouldn't necessarily care if it reads from file:// or s3:// > > to > > > load a > > > > > > > given > > > > > > > > > segment." > > > > > > > > > > > > > > > > > > Yes, this is what we are discussing in KIP. We are leaving > > the > > > details > > > > > > > of > > > > > > > > > loading segments to RLM read part instead of directly > > exposing > > > this in > > > > > > > the > > > > > > > > > Broker. This way we can keep the current Kafka code as it > is > > > without > > > > > > > > > changing the assumptions around the local disk. Let the RLM > > > handle the > > > > > > > > > remote storage part. > > > > > > > > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > Harsha > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Mon, Feb 4, 2019, at 12:54 PM, Ryanne Dolan wrote: > > > > > > > > > > Harsha, Sriharsha, Suresh, a couple thoughts: > > > > > > > > > > > > > > > > > > > > - How could this be used to leverage fast key-value > stores, > > > e.g. > > > > > > > > > Couchbase, > > > > > > > > > > which can serve individual records but maybe not entire > > > segments? Or > > > > > > > is > > > > > > > > > the > > > > > > > > > > idea to only support writing and fetching entire > segments? > > > Would it > > > > > > > make > > > > > > > > > > sense to support both? > > > > > > > > > > > > > > > > > > > > - Instead of defining a new interface and/or mechanism to > > > ETL segment > > > > > > > > > files > > > > > > > > > > from brokers to cold storage, can we just leverage Kafka > > > itself? In > > > > > > > > > > particular, we can already ETL records to HDFS via Kafka > > > Connect, > > > > > > > Gobblin > > > > > > > > > > etc -- we really just need a way for brokers to read > these > > > records > > > > > > > back. > > > > > > > > > > I'm wondering whether the new API could be limited to the > > > fetch, and > > > > > > > then > > > > > > > > > > existing ETL pipelines could be more easily leveraged. > For > > > example, > > > > > > > if > > > > > > > > > you > > > > > > > > > > already have an ETL pipeline from Kafka to HDFS, you > could > > > leave > > > > > > > that in > > > > > > > > > > place and just tell Kafka how to read these > > records/segments > > > from > > > > > > > cold > > > > > > > > > > storage when necessary. > > > > > > > > > > > > > > > > > > > > - I'm wondering if we could just add support for loading > > > segments > > > > > > > from > > > > > > > > > > remote URIs instead of from file, i.e. via plugins for > > > s3://, hdfs:// > > > > > > > > > etc. > > > > > > > > > > I suspect less broker logic would change in that case -- > > the > > > broker > > > > > > > > > > wouldn't necessarily care if it reads from file:// or > s3:// > > > to load a > > > > > > > > > given > > > > > > > > > > segment. > > > > > > > > > > > > > > > > > > > > Combining the previous two comments, I can imagine a URI > > > resolution > > > > > > > chain > > > > > > > > > > for segments. For example, first try > > > > > > > file:///logs/{topic}/{segment}.log, > > > > > > > > > > then s3://mybucket/{topic}/{date}/{segment}.log, etc, > > > leveraging your > > > > > > > > > > existing ETL pipeline(s). > > > > > > > > > > > > > > > > > > > > Ryanne > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Mon, Feb 4, 2019 at 12:01 PM Harsha <ka...@harsha.io> > > > wrote: > > > > > > > > > > > > > > > > > > > > > Hi All, > > > > > > > > > > > We are interested in adding tiered storage to Kafka. > More > > > > > > > > > details > > > > > > > > > > > about motivation and design are in the KIP. We are > > working > > > towards > > > > > > > an > > > > > > > > > > > initial POC. Any feedback or questions on this KIP are > > > welcome. > > > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > Harsha > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
