[Proposal] Add published segment cache in broker

[surekhasaharan]

Problem

Some of the sys.segments queries are slow, they are taking as long as ~10-20 sec, which is not desirable. The cause of this slowness is call from broker to coordinator API which happens every time a query is issued to sys.segments table, it’s the getMetaDataSegments (invokes coordinator api /druid/coordinator/v1/metadata/segments) method which gets called from the SegmentsTable#scan() in SystemSchema.java. Coordinator can potentially returns millions of segments and most of the time is spent in parsing the json response and creating DataSegment objects.

Motivation

It would be useful to make these queries faster as these are used in an interactive way by the end user today. In future a unified druid console can be built on top of sys tables(#6832) and the new segment locking can also benefit from all used segments present in broker.

Proposed Changes

To fix this performance bottleneck, plan to add :

1. segment cache in broker (phase 1)
2. a new api in coordinator (phase 2)

Phase 1

To speed up the sys.segments queries, in phase1 I want to add a published segments cache in broker. Broker already maintains a cache of all available segments via the BrokerServerView, as brokers are caching segments announced by historicals and realtime tasks, but not from metadata store (published segments are cached in coordinator only). It's going to be a pull model in this phase, where broker polls the coordinator to get latest published segments and updates it's cache periodically in a background thread.
Potential issue is it could lead to memory pressure on broker if the number of published segments is large. To minimize this memory pressure on Broker, the DataSegment instance should be shared between the “available” segments and “published” segments in broker. DataSegment already uses Interners, so we can keep interned DataSegment objects in the published and available segments, which prevents memory bloat by using the same reference in both places. Then, roughly, the extra memory would only be required for DataSegment objects which are “published but unavailable” segments, which ideally should be close to 0 segments.
Another point to consider is if a broker is configured to watch some particular tiers or datasources.
If these broker configs are enabled, then extra filtering would be done on watchedDataSources in the published segment cache in broker. There is no need to filter on watchedTiers because we can't tell the difference between published segments that are unavailable, and published segments that are served by tiers not in watchedTiers. This would be documented in the sys table docs.

Implementation details:

A new class MetadataSegmentView would be added which maintains the published segments cache in memory. The broker will keep a single DataSegment object in heap in case a segment overlaps in both published and available segment collection, via "interned" objects. BrokerServerView
maintains the available segments via DataSegment in ServerSelector, the AtomicReference for DataSegment object will store an interned DataSegment object. On deserialization of json stream of published segments from coordinator, each DataSegment object is interned with the same interner (used in ServerSelector) and cached in the published segments collection.
For the filter on watchedDataSources, if BrokerSegmentWatcherConfig#watchedDataSources is not null, then the segments with dataSource not in watchedDataSources would be filtered out and will not be stored in the published segment broker cache.

Phase 2

In phase 2 for this improvement, a more efficient coordinator API should be added. There can be several ways to add this new coordinator API, see rejected alternatives for other options considered.

This API returns a delta of added/removed segments and takes timestamp as argument. When broker comes up, it gets all the published segments from coordinator. Broker does following: orders the received segments by the timestamp (created_date), saves the published segment in it’s cache and keeps track of the last received segment’s timestamp. Subsequent calls to the coordinator api will only return the segments that have been added or removed since the last timestamp.The broker will poll the coordinator API at a regular interval to keep the published segment cache synced in a background thread. "added_ segments" delta can be computed based on the created_date, additional work would be required to compute the "deleted_segments" delta. Coordinator will need to maintain an in-memory list of deleted segments and will need to be notified when a segment gets killed external to coordinator (unless this behavior is changed as suggested in #6816). Since the deleted segments count can increase, to avoid memory pressure, coordinator can remember an hour(or some other configurable value) of deleted segments. In case, the requested timestamp is older than an hour, all the published segments can be resynced. In case of coordinator restart or leader change, again, it can send all the published segments.

New or Changed Public Interface

A new rest endpoint will be added to coordinator
GET /druid/coordinator/v1/metadata/segments/{timestamp}

Add a timestamp field to DataSegment object which represents the created_date from druid_segments table in metadata store.

Rejected Alternatives:

These options were also considered for the coordinator API

1. Coordinator sends just the ids of the published segments instead of complete DataSegment serialized objects, and then broker does a diff and finds out the segments which are not available, and then makes another call to get details for those segments. This approach was rejected because sometimes the segment_id list can be pretty large and it can cause a lot of network traffic between coordinator and broker processes and we may not achieve the performance improvement we are looking for.
2. Add a new table “druid_transactionlogs” to the metadata store, which keeps track of the segment addition and removal. The coordinator API can then query this table when it receives a GET request from broker for any timestamp, it can also query this to maintain it’s own cache. For example,

operation	segment_id	timestamp
add	s1	ts_0
disable	s2	ts_1
delete	s3	ts_1

It can use write ahead logging to take care of failures/restarts in any process. While this approach is good for maintaining consistency between coordinator and broker cache as well as fault tolerance, it may not give the speed improvement if we invoke the db call on each API invocation. Another challenge would be to keep the druid_segments table and druid_transactionlogs table in sync. Unless we need this for broader use cases, it may not be worth the extra design and effort.

[jihoonson]

@surekhasaharan thanks for raising this proposal! It would be really useful for #6319. It looks good to me overall, but I have some questions for details.

For phase 1,

• How does the broker keep available segments and published segments in memory? Are they stored separately or merged into a single data structure?

For phase 2,

• Killing segments is done by the kill tasks. How to notify the coordinator on killing segments? I suggested something related to this. Probably we can add TriggerAction which can be triggered by any metadata store change.
• Should we really need to modify dataSegment? I think it's probably fine to add a new type like DataSegmentWithCreatedDate unless createdDate needs to be used in many places in addition to the coordinator cache.
• If it's the coordinator API, it should start with /druid/coordinator/v1. Would you add the full URL for the new API?
• What is the return type of the new API? It should also be able to say "there's no delta for the given timestamp. I'm sending the full set of segments".

For both phase,

• Each broker can be configured to watch some particular tiers or dataSources. If these configurations are set, brokers don't have to retrieve all segments. How do the old & new APIs handle this?

[surekhasaharan]

@jihoonson Thanks for reading the proposal and your comments. I have added more details in the proposal for phase1 to address your questions.

[jihoonson]

@surekhasaharan thanks. I have some follow-up questions.

If these broker configs are enabled, then extra filtering would be done on watchedDataSources in the published segment cache in broker.

This part looks vague. Would you please make it clear what configurations you're referring? Also please add why we don't have to consider tiers.

One caveat with this is, the new MetadataSegmentView has a dependency that BrokerServerView should be initialized before MetadataSegmentView starts caching published segments.

Since segments are always published and then loaded by historicals, segments should appear first in the metadata store. How do you reuse the same instance if a segment was only in MetadataSegmentView is added to brokerServerView?

[surekhasaharan]

@jihoonson thanks. Added more details on broker config to make it clearer.

Since segments are always published and then loaded by historicals, segments should appear first in the metadata store. How do you reuse the same instance if a segment was only in MetadataSegmentView is added to brokerServerView?

Changed the design for this part, to use the interned objects, this approach does not depend on the order in which DataSegment objects appear in broker heap.

[gianm]

It looks like phase 1 was implemented by #6901, and phase 2 remains unimplemented.

[leventov]

@surekhasaharan did you consider the following designs:

• Brokers delegate Druid SQL queries against sys.segments to Coordinator. So brokers don't have to deal with published (used) segments in their memory at all.
• Brokers access the metadata store to update their view of published (used) segments directly, without Coordinator as a proxy.
• For memory savings, adding an extra flags to DataSegment object: isAvailable, isPublished, and using a single unified ConcurrentHashMap view for the purposes of both BrokerServerView and MetadataSegmentView.

What do you think about these?

[gianm]

Brokers delegate Druid SQL queries against sys.segments to Coordinator. So brokers don't have to deal with published (used) segments in their memory at all.

We had talked a bit about this idea offline last week and I believe generally thought it was a good idea in isolation. In particular, we had thought about moving the entire sys schema implementation to the Coordinator and having the Broker send any SQL queries on sys over there. Users could also query sys tables on the Coordinator directly if they wanted.

However, there were two counterpoints raised:

• It is a somewhat common request from our users to add an option to the Broker to either fail a query, or provide a warning, if that query is returning partial results because the underlying segments are partially unavailable. Right now, if some segments are not available, the Broker's query runner does not know (it has zero information about unavailable segments). So it will simply return partial results, without informing the caller. Definitely you could imagine life being better if the caller knew about this. For the feature to work properly the Broker needs to know about segments that are published but unavailable, which is the same
• [Proposal] Minimizing the lock granularity for indexing #6319 contemplates a design for finer-grained locking and finer-grained atomic overwriting that would be useful for a compaction feature. Its design, as currently contemplated, requires the Broker to retain information about segments that are published but unavailable.

Brokers access the metadata store to update their view of published (used) segments directly, without Coordinator as a proxy.

This one was considered as well, the main counterpoint raised was that currently, only the Coordinator and Overlord have a direct dependency on the metadata store, and there was a desire to keep it that way. In turn there were two reasons for that:

• A desire to minimize the number of services that depend on the metadata store, both for architectural complexity and load reasons.
• A thought that someday, in the future, we might want to bring the metadata store 'in house' and have it be something the Coordinator does on its own without an external dependency. This could be through implementing Raft or some other consensus protocol on the Coordinators, and using that to allow the Coordinators to take on the functionality of both ZooKeeper and the metadata store. The motivation would be to simplify deployment and reduce the number of mandatory external dependencies. In this case, any other server type that needs to interact with metadata would have to do so through Coordinator APIs. So we might as well stick to that now to make a potential future change easier.

For memory savings, adding an extra flags to DataSegment object: isAvailable, isPublished, and using a single unified ConcurrentHashMap view for the purposes of both BrokerServerView and MetadataSegmentView.

This was not discussed. I haven't thought through the implications fully but I would worry that it means either DataSegment must no longer be immutable (because isAvailable, etc can change) or that it remains immutable but can no longer be interned effectively (because there's no longer a single DataSegment object per SegmentId). So I would look at those things first when evaluating this idea.

[leventov]

A desire to minimize the number of services that depend on the metadata store, both for architectural complexity and load reasons.

It might be (sometimes, or even often) the opposite of that:

• Proxying through Coordinator makes the latter a more important node. If the Coordinator is down, the data on Brokers becomes stale, or newly started Brokers cannot load data until the Coordinator is up. Metadata store itself is generally considered a much more stable piece, it's expected to be down more rarely than Coordinators. Generally, the promise of Druid is that whichever type of node is down, the rest of the cluster can continue to operate somewhat normally.
• Metadata store is a database service managed in the cloud (such as Amazon Aurora) that is more scalable than a single Coordinator node.

Also, it would allow simplifying the codebase, because there would be a single abstraction used in both Coordinator and Brokers. I initially turned my attention to this proposal when I was forced to write explanations of how MetadataSegmentView is different from SegmentsMetadata.

[leventov]

Another design question, regarding #6901: was it considered to not add druid.sql.planner.metadataSegmentCacheEnable configuration parameter? Perhaps, just start caching after the first access to sys.segments?