keep track of first offset and timestamp per segment (#817)

Saves a hash with the first offset/timestamp for each segment, greatly reducing the time (and memory used) to find a message by offset and timestamps.

---------

Co-authored-by: Carl Hörberg <carl@84codes.com>

CHANGELOG.md

@@ -18,6 +18,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Remove the 'reset vhost' feature [#822](https://github.com/cloudamqp/lavinmq/pull/822/files)
 
+### Added
+
+- Added some indexing for streams, greatly increasing performance when looking up by offset or timestamp [#817](https://github.com/cloudamqp/lavinmq/pull/817)
+
 ## [2.0.0] - 2024-10-31
 
 With the release of 2.0.0 we introduce High Availablility for LavinMQ in the form of clustering. With clustering, LavinMQ replicates data between nodes with our own replication protocol, and uses etcd for leader election. See [this post](https://lavinmq.com/blog/lavinmq-high-availability) in the LavinMQ blog or the [readme](https://github.com/cloudamqp/lavinmq?tab=readme-ov-file#clustering) for more information about clustering.

src/lavinmq/amqp/queue/stream_queue_message_store.cr

@@ -9,10 +9,13 @@ module LavinMQ::AMQP
       property max_age : Time::Span | Time::MonthSpan | Nil
       getter last_offset : Int64
       @segment_last_ts = Hash(UInt32, Int64).new(0i64) # used for max-age
+      @offset_index = Hash(UInt32, Int64).new          # segment_id => offset of first msg
+      @timestamp_index = Hash(UInt32, Int64).new       # segment_id => ts of first msg
 
       def initialize(*args, **kwargs)
         super
         @last_offset = get_last_offset
+        build_segment_indexes
         drop_overflow
       end
 
@@ -62,7 +65,7 @@ module LavinMQ::AMQP
       end
 
       private def find_offset_in_segments(offset : Int | Time) : Tuple(Int64, UInt32, UInt32)
-        segment = @segments.first_key
+        segment = offset_index_lookup(offset)
         pos = 4u32
         msg_offset = 0i64
         loop do
@@ -88,6 +91,23 @@ module LavinMQ::AMQP
         {msg_offset, segment, pos}
       end
 
+      private def offset_index_lookup(offset) : UInt32
+        seg = @segments.first_key
+        case offset
+        when Int
+          @offset_index.each do |seg_id, first_seg_offset|
+            break if first_seg_offset > offset
+            seg = seg_id
+          end
+        when Time
+          @timestamp_index.each do |seg_id, first_seg_ts|
+            break if Time.unix_ms(first_seg_ts) > offset
+            seg = seg_id
+          end
+        end
+        seg
+      end
+
       def shift?(consumer : AMQP::StreamConsumer) : Envelope?
         raise ClosedError.new if @closed
 
@@ -147,6 +167,8 @@ module LavinMQ::AMQP
       private def open_new_segment(next_msg_size = 0) : MFile
         super.tap do
           drop_overflow
+          @offset_index[@segments.last_key] = @last_offset + 1
+          @timestamp_index[@segments.last_key] = RoughTime.unix_ms
         end
       end
 
@@ -178,6 +200,8 @@ module LavinMQ::AMQP
           msg_count = @segment_msg_count.delete(seg_id)
           @size -= msg_count if msg_count
           @segment_last_ts.delete(seg_id)
+          @offset_index.delete(seg_id)
+          @timestamp_index.delete(seg_id)
           @bytesize -= mfile.size - 4
           mfile.delete.close
           @replicator.try &.delete_file(mfile.path)
@@ -211,6 +235,17 @@ module LavinMQ::AMQP
       private def offset_from_headers(headers) : Int64
         headers.not_nil!("Message lacks headers")["x-stream-offset"].as(Int64)
       end
+
+      private def build_segment_indexes
+        @segments.each do |seg_id, mfile|
+          msg = BytesMessage.from_bytes(mfile.to_slice + 4u32)
+          @offset_index[seg_id] = offset_from_headers(msg.properties.headers)
+          @timestamp_index[seg_id] = msg.timestamp
+        rescue IndexError
+          @offset_index[seg_id] = @last_offset
+          @timestamp_index[seg_id] = RoughTime.unix_ms
+        end
+      end
     end
   end
 end