Step nine of the failover runbook
said this.
Do not restart the replica
until replication lag reads zero.
You had read it before.
It was late,
you knew roughly what it said,
and you read it fast.
You restarted the replica.
Lag was at forty seconds.
Here is the mechanism.
Reading feels like a smooth glide
along the line.
It is not.
Your eyes jump.
Each jump is a saccade,
and you see almost nothing during it.
Between jumps
the eyes stop for a fixation,
usually around a quarter of a second,
and that is when the reading happens.
Even then,
sharp vision covers a tiny area,
a few letters either side
of where you are looking.
The rest of the line
is a blur that tells you
roughly where the next word is
and how long it is.
So readers do not look at every word.
Short words
are skipped far more often
than long ones,
and identified,
usually correctly,
from the edge of the previous stop.
Usually.
The word not
is three letters long.
Read faster,
and you make longer jumps,
fewer stops,
and more guesses.
There is something else
you give up.
In ordinary reading,
roughly one eye movement in ten
goes backwards.
You barely notice them.
They are where you catch a mistake
and repair the sentence.
Skimming squeezes them out.
A large review of reading research
by Keith Rayner and colleagues,
published in 2016,
found no trick
that lets people read much faster
than normal
without understanding less.
Skimming is fine
for finding the right page.
It is the wrong way
to follow it.
So change speed
when the text changes job.
Anything that tells you
what to type into production
gets read at normal speed,
one step at a time,
with your cursor on the line.
Read each step aloud,
quietly,
before you do it.
Speaking forces the eye
to land on every word.
And fix the runbook.
Write steps
that say what to do,
not what to avoid.
Wait until replication lag reads zero.
Then restart the replica.
An instruction
should still be correct
when somebody tired
reads only the long words.
– Serguey Asael Shinder
