Based on the questions and suggestions raised by Andrey Yegorov, I have updated my proposal. Here is the revised introduction:
The new features will not have any impact on the existing architecture implementation. We have introduced a new implementation called DirectDbLedgerStorage, which eliminates the use of journal pre-writing logs. Now, ledger data can be directly written to the SSD disk, eliminating the need for data to be written twice to the SSD disk. Data in the SSD disk cache is evicted to the HDD disk based on the write time at the granularity of entry logs. Furthermore, we have added support for write degradation. When the SSD disk cache reaches the warning threshold, the system automatically switches to the journal+ledger approach to ensure system stability. On 2023/05/29 08:49:37 Gavin Gao wrote: > The answer for question 1 is `yes`. > After reading through it, your main oncerning is the impact on journal write > throughput. I need to do some load test to verfiy How much impact is added? > > For question 2: Other solutions like linux's storage tiering the drawback is: > 1. Both FlashCache/OpenCAS, in every mode, will flush data back to the SSD > cache, similar to PageCache, causing cache pollution. > 2.Cache Miss results in an additional access to the device, increasing > latency. > 3. All metadata is maintained by the operating system, leading to increased > memory consumption by the kernel. > > On 2023/05/26 17:13:13 Andrey Yegorov wrote: > > Semi-random notes: > > > > 1. Are you planning to use the journal's SSD as cache? > > > > I guess this is a typical thing that pops up after looking at e.g. 256GB > > SSD with e.g. only 10GB actually utilized by journals. > > The problem with that idea is that reads and writes start mixing up on the > > same disk. > > The value of the dedicated disk for journal (even if that's HDD) is that it > > only gets sequential write i/o. > > As soon as you add cache writes (and random reads, you are adding cache to > > read from it) you'll see a significant drop in write throughput and > > significant increase in latency for the journal, fsync latency spikes up. I > > do not have access to the numbers from the tests I've done a few years ago > > but that was the conclusion. > > > > At the end you are sacrificing write performance for read performance which > > is ok for some usecases but not for all. > > You can simulate this by running anything that generates > > writes/reads/deletes to the journal's ssd (at expected cache rates) during > > your regular load tests. You won't see read improvements but you will see > > the write perf changes. > > > > 2. In case you have dedicated SSD for the cache on the hardware that you > > manage I'd look at other options first, like linux's storage tiering > > https://blog.delouw.ch/2020/01/29/using-lvm-cache-for-storage-tiering/ and > > proceed after having the perf test results. > > > > 3. Writing to cache as an entry log and then moving the entry log to a > > slower class of storage (as I understood; also this way the index can be > > reused) complicates compaction. > > It also assumes that journal(s) keep enough data for recovery of cached > > data in case cache is lost. Both compaction and recovery from journals are > > getting affected by this. > > > > 4. for cloud-based deployments using local ssds as read cache could be an > > option as long as we do not expect these disks to be reliable. > > This means data goes to entry logs as usual and to the cache that can be > > recreated with newly written data (not rebuilt) > > One could use e.g. RocksDB and entries with TTL (is there a size-based > > expiration in rocksdb?) > > This has a few gotchas of its own: > > - compaction will have to remove data from cache > > - RocksDB's internal compaction can introduce latency spikes (it can block > > reads/writes IIRC). > > - to avoid returning stale data from the cache we need to read index > > first to confirm the entry exists, then try cache, fallback to ledger > > storage > > As RocksDB alternative in this case, ehcache is mature, apache 2 licensed, > > and supports tiering/cache persistence but I do not have any experience > > with it. https://www.ehcache.org/documentation/3.10/tiering.html > > > > 5. Cache has to be abstracted as an interface and support > > pluggable/configurable implementations. the abstracted interface has to > > define expected side-effects on compaction/recovery, how to access it > > to prevent reading stale (compacted out) data. > > > > To be clear, I am ok with the idea of having an optional tiered > > storage/entry cache but I do not think that proposed implementation is the > > way to go. > > > > > > On Thu, May 25, 2023 at 7:37 PM Gavin gao <gaozhang...@gmail.com> wrote: > > > > > In a typical bookkeeper deployment, SSD disks are used to store Journal > > > log > > > data, while HDD disks are used to store Ledger data. Data writes are > > > initially stored in memory and then asynchronously flushed to the HDD disk > > > in the background. However, due to memory limitations, the amount of data > > > that can be cached is restricted. Consequently, requests for historical > > > data ultimately rely on the HDD disk, which becomes a bottleneck for the > > > entire Bookkeeper cluster. Moreover, during data recovery processes > > > following node failures, a substantial amount of historical data needs to > > > be read from the HDD disk, leading to the disk's I/O utilization reaching > > > maximum capacity and resulting in significant read request delays or > > > failures. > > > > > > To address these challenges, a new architecture is proposed: the > > > introduction of a disk cache between the memory cache and the HDD disk, > > > utilizing an SSD disk as an intermediary medium to significantly extend > > > data caching duration. The data flow is as follows: journal -> write cache > > > -> SSD cache -> HDD disk. The SSD disk cache functions as a regular > > > LedgerStorage layer and is compatible with all existing LedgerStorage > > > implementations. The following outlines the process: > > > > > > 1. Data eviction from the disk cache to the Ledger data disk occurs on > > > a > > > per-log file basis. > > > 2. A new configuration parameter, diskCacheRetentionTime, is added to > > > set the duration for which hot data is retained. Files with write > > > timestamps older than the retention time will be evicted to the Ledger > > > data > > > disk. > > > 3. A new configuration parameter, diskCacheThreshold, is added. If the > > > disk cache utilization exceeds the threshold, the eviction process is > > > accelerated. Data is evicted to the Ledger data disk based on the order > > > of > > > file writes until the disk space recovers above the threshold. > > > 4. A new thread, ColdStorageArchiveThread, is introduced to > > > periodically > > > evict data from the disk cache to the Ledger data disk. > > > > > > > > > -- > > Andrey Yegorov > > >
