I don't know if this is useful for anyone else but I have been running a
quorum sidecar alongside my servers:
https://github.com/phrocker/veculo/blob/main/server/wal-quorum-sidecar/ARCHITECTURE.md
in conjunction with a specialized GCS caching file system ( using local
nvme pvcs ), I can achieve low latency writes. With this approach on this
cluster w/ 56k queries I averaged 288ms for lookups. +r and !0 are in GCS.
The code is all in the aforementioned repo save for the deployment code in
a private repo. Hope it spawns some ideas...

[image: image.png]


On Thu, Apr 16, 2026 at 2:59 PM Keith Turner <[email protected]> wrote:

> The advice to store the root and metadata table was based on the fact that
> the root tablet used to rely on DFS file listing to know what files it
> had.  S3 file listings are eventually consistent and not reliable so that
> could lead to the root tablet missing data.  In Accumulo 2.1 the list of
> root tablet files is stored in zookeeper and it no longer relies on DFS
> files listing operations to know what files it has. With this change in 2.1
> it may be safe to store root and metadata files in S3.  However tablet
> servers write ahead logs probably still need to be stored in a DFS impl
> that supports reliable sync operations, S3 may not support this.
>
> On Tue, Apr 14, 2026 at 9:13 PM McClure, Bruce MR 2 via user <
> [email protected]> wrote:
>
>> *UNOFFICIAL*
>>
>> Hi,
>>
>>
>>
>> I am just wondering if there is any update since the advice in 2019 that
>> said Accumulo can run on S3 as long as the root and metadata tables are in
>> HDFS?
>>
>>
>>
>> Has any performance benchmarking been done comparing Accumulo running on
>> S3 versus HDFS?  (for comparable hardware/networking configuration).
>>
>>
>>
>> Thank you,
>>
>>
>>
>> Bruce.
>>
>>
>>
>

Reply via email to